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 simmilar 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         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1157         memset(opt, 0xff, option_len);
1158         opt->length = data_len / 4;
1159         opt->r1 = 0;
1160         opt->r2 = 0;
1161         opt->r3 = 0;
1162
1163         /* If no mask has been prodived we assume an exact match. */
1164         if (!depth)
1165                 return sizeof(struct geneve_opt) + data_len;
1166
1167         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1168                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1169                 return -EINVAL;
1170         }
1171
1172         err = nla_parse_nested_deprecated(tb,
1173                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1174                                           nla, geneve_opt_policy, extack);
1175         if (err < 0)
1176                 return err;
1177
1178         /* We are not allowed to omit any of CLASS, TYPE or DATA
1179          * fields from the key.
1180          */
1181         if (!option_len &&
1182             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1183              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1184              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1185                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1186                 return -EINVAL;
1187         }
1188
1189         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1190          * for the mask.
1191          */
1192         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1193                 int new_len = key->enc_opts.len;
1194
1195                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1196                 data_len = nla_len(data);
1197                 if (data_len < 4) {
1198                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1199                         return -ERANGE;
1200                 }
1201                 if (data_len % 4) {
1202                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1203                         return -ERANGE;
1204                 }
1205
1206                 new_len += sizeof(struct geneve_opt) + data_len;
1207                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1208                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1209                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1210                         return -ERANGE;
1211                 }
1212                 opt->length = data_len / 4;
1213                 memcpy(opt->opt_data, nla_data(data), data_len);
1214         }
1215
1216         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1217                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1218                 opt->opt_class = nla_get_be16(class);
1219         }
1220
1221         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1222                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1223                 opt->type = nla_get_u8(type);
1224         }
1225
1226         return sizeof(struct geneve_opt) + data_len;
1227 }
1228
1229 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1230                             int depth, int option_len,
1231                             struct netlink_ext_ack *extack)
1232 {
1233         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1234         struct vxlan_metadata *md;
1235         int err;
1236
1237         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1238         memset(md, 0xff, sizeof(*md));
1239
1240         if (!depth)
1241                 return sizeof(*md);
1242
1243         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1244                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1245                 return -EINVAL;
1246         }
1247
1248         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1249                                vxlan_opt_policy, extack);
1250         if (err < 0)
1251                 return err;
1252
1253         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1254                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1255                 return -EINVAL;
1256         }
1257
1258         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1259                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1260                 md->gbp &= VXLAN_GBP_MASK;
1261         }
1262
1263         return sizeof(*md);
1264 }
1265
1266 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1267                              int depth, int option_len,
1268                              struct netlink_ext_ack *extack)
1269 {
1270         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1271         struct erspan_metadata *md;
1272         int err;
1273
1274         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1275         memset(md, 0xff, sizeof(*md));
1276         md->version = 1;
1277
1278         if (!depth)
1279                 return sizeof(*md);
1280
1281         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1282                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1283                 return -EINVAL;
1284         }
1285
1286         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1287                                erspan_opt_policy, extack);
1288         if (err < 0)
1289                 return err;
1290
1291         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1292                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1293                 return -EINVAL;
1294         }
1295
1296         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1297                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1298
1299         if (md->version == 1) {
1300                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1301                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1302                         return -EINVAL;
1303                 }
1304                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1305                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1306                         memset(&md->u, 0x00, sizeof(md->u));
1307                         md->u.index = nla_get_be32(nla);
1308                 }
1309         } else if (md->version == 2) {
1310                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1311                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1312                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1313                         return -EINVAL;
1314                 }
1315                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1316                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1317                         md->u.md2.dir = nla_get_u8(nla);
1318                 }
1319                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1320                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1321                         set_hwid(&md->u.md2, nla_get_u8(nla));
1322                 }
1323         } else {
1324                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1325                 return -EINVAL;
1326         }
1327
1328         return sizeof(*md);
1329 }
1330
1331 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1332                           int depth, int option_len,
1333                           struct netlink_ext_ack *extack)
1334 {
1335         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1336         struct gtp_pdu_session_info *sinfo;
1337         u8 len = key->enc_opts.len;
1338         int err;
1339
1340         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1341         memset(sinfo, 0xff, option_len);
1342
1343         if (!depth)
1344                 return sizeof(*sinfo);
1345
1346         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1347                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1348                 return -EINVAL;
1349         }
1350
1351         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1352                                gtp_opt_policy, extack);
1353         if (err < 0)
1354                 return err;
1355
1356         if (!option_len &&
1357             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1358              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1359                 NL_SET_ERR_MSG_MOD(extack,
1360                                    "Missing tunnel key gtp option pdu type or qfi");
1361                 return -EINVAL;
1362         }
1363
1364         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1365                 sinfo->pdu_type =
1366                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1367
1368         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1369                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1370
1371         return sizeof(*sinfo);
1372 }
1373
1374 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1375                           struct fl_flow_key *mask,
1376                           struct netlink_ext_ack *extack)
1377 {
1378         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1379         int err, option_len, key_depth, msk_depth = 0;
1380
1381         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1382                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1383                                              enc_opts_policy, extack);
1384         if (err)
1385                 return err;
1386
1387         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1388
1389         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1390                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1391                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1392                                                      enc_opts_policy, extack);
1393                 if (err)
1394                         return err;
1395
1396                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1397                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1398                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1399                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1400                         return -EINVAL;
1401                 }
1402         }
1403
1404         nla_for_each_attr(nla_opt_key, nla_enc_key,
1405                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1406                 switch (nla_type(nla_opt_key)) {
1407                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1408                         if (key->enc_opts.dst_opt_type &&
1409                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1410                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1411                                 return -EINVAL;
1412                         }
1413                         option_len = 0;
1414                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1415                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1416                                                        key_depth, option_len,
1417                                                        extack);
1418                         if (option_len < 0)
1419                                 return option_len;
1420
1421                         key->enc_opts.len += option_len;
1422                         /* At the same time we need to parse through the mask
1423                          * in order to verify exact and mask attribute lengths.
1424                          */
1425                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1426                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1427                                                        msk_depth, option_len,
1428                                                        extack);
1429                         if (option_len < 0)
1430                                 return option_len;
1431
1432                         mask->enc_opts.len += option_len;
1433                         if (key->enc_opts.len != mask->enc_opts.len) {
1434                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1435                                 return -EINVAL;
1436                         }
1437                         break;
1438                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1439                         if (key->enc_opts.dst_opt_type) {
1440                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1441                                 return -EINVAL;
1442                         }
1443                         option_len = 0;
1444                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1445                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1446                                                       key_depth, option_len,
1447                                                       extack);
1448                         if (option_len < 0)
1449                                 return option_len;
1450
1451                         key->enc_opts.len += option_len;
1452                         /* At the same time we need to parse through the mask
1453                          * in order to verify exact and mask attribute lengths.
1454                          */
1455                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1456                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1457                                                       msk_depth, option_len,
1458                                                       extack);
1459                         if (option_len < 0)
1460                                 return option_len;
1461
1462                         mask->enc_opts.len += option_len;
1463                         if (key->enc_opts.len != mask->enc_opts.len) {
1464                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1465                                 return -EINVAL;
1466                         }
1467                         break;
1468                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1469                         if (key->enc_opts.dst_opt_type) {
1470                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1471                                 return -EINVAL;
1472                         }
1473                         option_len = 0;
1474                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1475                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1476                                                        key_depth, option_len,
1477                                                        extack);
1478                         if (option_len < 0)
1479                                 return option_len;
1480
1481                         key->enc_opts.len += option_len;
1482                         /* At the same time we need to parse through the mask
1483                          * in order to verify exact and mask attribute lengths.
1484                          */
1485                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1486                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1487                                                        msk_depth, option_len,
1488                                                        extack);
1489                         if (option_len < 0)
1490                                 return option_len;
1491
1492                         mask->enc_opts.len += option_len;
1493                         if (key->enc_opts.len != mask->enc_opts.len) {
1494                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1495                                 return -EINVAL;
1496                         }
1497                         break;
1498                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1499                         if (key->enc_opts.dst_opt_type) {
1500                                 NL_SET_ERR_MSG_MOD(extack,
1501                                                    "Duplicate type for gtp options");
1502                                 return -EINVAL;
1503                         }
1504                         option_len = 0;
1505                         key->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1506                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1507                                                     key_depth, option_len,
1508                                                     extack);
1509                         if (option_len < 0)
1510                                 return option_len;
1511
1512                         key->enc_opts.len += option_len;
1513                         /* At the same time we need to parse through the mask
1514                          * in order to verify exact and mask attribute lengths.
1515                          */
1516                         mask->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1517                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1518                                                     msk_depth, option_len,
1519                                                     extack);
1520                         if (option_len < 0)
1521                                 return option_len;
1522
1523                         mask->enc_opts.len += option_len;
1524                         if (key->enc_opts.len != mask->enc_opts.len) {
1525                                 NL_SET_ERR_MSG_MOD(extack,
1526                                                    "Key and mask miss aligned");
1527                                 return -EINVAL;
1528                         }
1529                         break;
1530                 default:
1531                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1532                         return -EINVAL;
1533                 }
1534
1535                 if (!msk_depth)
1536                         continue;
1537
1538                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1539                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1540                         return -EINVAL;
1541                 }
1542                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1543         }
1544
1545         return 0;
1546 }
1547
1548 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1549                                 struct netlink_ext_ack *extack)
1550 {
1551         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1552                 NL_SET_ERR_MSG_ATTR(extack, tb,
1553                                     "no trk, so no other flag can be set");
1554                 return -EINVAL;
1555         }
1556
1557         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1558             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1559                 NL_SET_ERR_MSG_ATTR(extack, tb,
1560                                     "new and est are mutually exclusive");
1561                 return -EINVAL;
1562         }
1563
1564         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1565             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1566                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1567                 NL_SET_ERR_MSG_ATTR(extack, tb,
1568                                     "when inv is set, only trk may be set");
1569                 return -EINVAL;
1570         }
1571
1572         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1573             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1574                 NL_SET_ERR_MSG_ATTR(extack, tb,
1575                                     "new and rpl are mutually exclusive");
1576                 return -EINVAL;
1577         }
1578
1579         return 0;
1580 }
1581
1582 static int fl_set_key_ct(struct nlattr **tb,
1583                          struct flow_dissector_key_ct *key,
1584                          struct flow_dissector_key_ct *mask,
1585                          struct netlink_ext_ack *extack)
1586 {
1587         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1588                 int err;
1589
1590                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1591                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1592                         return -EOPNOTSUPP;
1593                 }
1594                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1595                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1596                                sizeof(key->ct_state));
1597
1598                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1599                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1600                                            extack);
1601                 if (err)
1602                         return err;
1603
1604         }
1605         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1606                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1607                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1608                         return -EOPNOTSUPP;
1609                 }
1610                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1611                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1612                                sizeof(key->ct_zone));
1613         }
1614         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1615                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1616                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1617                         return -EOPNOTSUPP;
1618                 }
1619                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1620                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1621                                sizeof(key->ct_mark));
1622         }
1623         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1624                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1625                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1626                         return -EOPNOTSUPP;
1627                 }
1628                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1629                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1630                                sizeof(key->ct_labels));
1631         }
1632
1633         return 0;
1634 }
1635
1636 static bool is_vlan_key(struct nlattr *tb, __be16 *ethertype,
1637                         struct fl_flow_key *key, struct fl_flow_key *mask,
1638                         int vthresh)
1639 {
1640         const bool good_num_of_vlans = key->num_of_vlans.num_of_vlans > vthresh;
1641
1642         if (!tb) {
1643                 *ethertype = 0;
1644                 return good_num_of_vlans;
1645         }
1646
1647         *ethertype = nla_get_be16(tb);
1648         if (good_num_of_vlans || eth_type_vlan(*ethertype))
1649                 return true;
1650
1651         key->basic.n_proto = *ethertype;
1652         mask->basic.n_proto = cpu_to_be16(~0);
1653         return false;
1654 }
1655
1656 static int fl_set_key(struct net *net, struct nlattr **tb,
1657                       struct fl_flow_key *key, struct fl_flow_key *mask,
1658                       struct netlink_ext_ack *extack)
1659 {
1660         __be16 ethertype;
1661         int ret = 0;
1662
1663         if (tb[TCA_FLOWER_INDEV]) {
1664                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1665                 if (err < 0)
1666                         return err;
1667                 key->meta.ingress_ifindex = err;
1668                 mask->meta.ingress_ifindex = 0xffffffff;
1669         }
1670
1671         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1672                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1673                        sizeof(key->eth.dst));
1674         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1675                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1676                        sizeof(key->eth.src));
1677         fl_set_key_val(tb, &key->num_of_vlans,
1678                        TCA_FLOWER_KEY_NUM_OF_VLANS,
1679                        &mask->num_of_vlans,
1680                        TCA_FLOWER_UNSPEC,
1681                        sizeof(key->num_of_vlans));
1682
1683         if (is_vlan_key(tb[TCA_FLOWER_KEY_ETH_TYPE], &ethertype, key, mask, 0)) {
1684                 fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1685                                 TCA_FLOWER_KEY_VLAN_PRIO,
1686                                 TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1687                                 &key->vlan, &mask->vlan);
1688
1689                 if (is_vlan_key(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE],
1690                                 &ethertype, key, mask, 1)) {
1691                         fl_set_key_vlan(tb, ethertype,
1692                                         TCA_FLOWER_KEY_CVLAN_ID,
1693                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1694                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1695                                         &key->cvlan, &mask->cvlan);
1696                         fl_set_key_val(tb, &key->basic.n_proto,
1697                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1698                                        &mask->basic.n_proto,
1699                                        TCA_FLOWER_UNSPEC,
1700                                        sizeof(key->basic.n_proto));
1701                 }
1702         }
1703
1704         if (key->basic.n_proto == htons(ETH_P_PPP_SES))
1705                 fl_set_key_pppoe(tb, &key->pppoe, &mask->pppoe, key, mask);
1706
1707         if (key->basic.n_proto == htons(ETH_P_IP) ||
1708             key->basic.n_proto == htons(ETH_P_IPV6)) {
1709                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1710                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1711                                sizeof(key->basic.ip_proto));
1712                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1713         }
1714
1715         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1716                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1717                 mask->control.addr_type = ~0;
1718                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1719                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1720                                sizeof(key->ipv4.src));
1721                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1722                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1723                                sizeof(key->ipv4.dst));
1724         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1725                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1726                 mask->control.addr_type = ~0;
1727                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1728                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1729                                sizeof(key->ipv6.src));
1730                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1731                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1732                                sizeof(key->ipv6.dst));
1733         }
1734
1735         if (key->basic.ip_proto == IPPROTO_TCP) {
1736                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1737                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1738                                sizeof(key->tp.src));
1739                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1740                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1741                                sizeof(key->tp.dst));
1742                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1743                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1744                                sizeof(key->tcp.flags));
1745         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1746                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1747                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1748                                sizeof(key->tp.src));
1749                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1750                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1751                                sizeof(key->tp.dst));
1752         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1753                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1754                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1755                                sizeof(key->tp.src));
1756                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1757                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1758                                sizeof(key->tp.dst));
1759         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1760                    key->basic.ip_proto == IPPROTO_ICMP) {
1761                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1762                                &mask->icmp.type,
1763                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1764                                sizeof(key->icmp.type));
1765                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1766                                &mask->icmp.code,
1767                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1768                                sizeof(key->icmp.code));
1769         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1770                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1771                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1772                                &mask->icmp.type,
1773                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1774                                sizeof(key->icmp.type));
1775                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1776                                &mask->icmp.code,
1777                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1778                                sizeof(key->icmp.code));
1779         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1780                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1781                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1782                 if (ret)
1783                         return ret;
1784         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1785                    key->basic.n_proto == htons(ETH_P_RARP)) {
1786                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1787                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1788                                sizeof(key->arp.sip));
1789                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1790                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1791                                sizeof(key->arp.tip));
1792                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1793                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1794                                sizeof(key->arp.op));
1795                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1796                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1797                                sizeof(key->arp.sha));
1798                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1799                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1800                                sizeof(key->arp.tha));
1801         } else if (key->basic.ip_proto == IPPROTO_L2TP) {
1802                 fl_set_key_val(tb, &key->l2tpv3.session_id,
1803                                TCA_FLOWER_KEY_L2TPV3_SID,
1804                                &mask->l2tpv3.session_id, TCA_FLOWER_UNSPEC,
1805                                sizeof(key->l2tpv3.session_id));
1806         }
1807
1808         if (key->basic.ip_proto == IPPROTO_TCP ||
1809             key->basic.ip_proto == IPPROTO_UDP ||
1810             key->basic.ip_proto == IPPROTO_SCTP) {
1811                 ret = fl_set_key_port_range(tb, key, mask, extack);
1812                 if (ret)
1813                         return ret;
1814         }
1815
1816         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1817             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1818                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1819                 mask->enc_control.addr_type = ~0;
1820                 fl_set_key_val(tb, &key->enc_ipv4.src,
1821                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1822                                &mask->enc_ipv4.src,
1823                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1824                                sizeof(key->enc_ipv4.src));
1825                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1826                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1827                                &mask->enc_ipv4.dst,
1828                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1829                                sizeof(key->enc_ipv4.dst));
1830         }
1831
1832         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1833             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1834                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1835                 mask->enc_control.addr_type = ~0;
1836                 fl_set_key_val(tb, &key->enc_ipv6.src,
1837                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1838                                &mask->enc_ipv6.src,
1839                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1840                                sizeof(key->enc_ipv6.src));
1841                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1842                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1843                                &mask->enc_ipv6.dst,
1844                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1845                                sizeof(key->enc_ipv6.dst));
1846         }
1847
1848         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1849                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1850                        sizeof(key->enc_key_id.keyid));
1851
1852         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1853                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1854                        sizeof(key->enc_tp.src));
1855
1856         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1857                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1858                        sizeof(key->enc_tp.dst));
1859
1860         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1861
1862         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1863                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1864                        sizeof(key->hash.hash));
1865
1866         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1867                 ret = fl_set_enc_opt(tb, key, mask, extack);
1868                 if (ret)
1869                         return ret;
1870         }
1871
1872         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1873         if (ret)
1874                 return ret;
1875
1876         if (tb[TCA_FLOWER_KEY_FLAGS])
1877                 ret = fl_set_key_flags(tb, &key->control.flags,
1878                                        &mask->control.flags, extack);
1879
1880         return ret;
1881 }
1882
1883 static void fl_mask_copy(struct fl_flow_mask *dst,
1884                          struct fl_flow_mask *src)
1885 {
1886         const void *psrc = fl_key_get_start(&src->key, src);
1887         void *pdst = fl_key_get_start(&dst->key, src);
1888
1889         memcpy(pdst, psrc, fl_mask_range(src));
1890         dst->range = src->range;
1891 }
1892
1893 static const struct rhashtable_params fl_ht_params = {
1894         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1895         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1896         .automatic_shrinking = true,
1897 };
1898
1899 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1900 {
1901         mask->filter_ht_params = fl_ht_params;
1902         mask->filter_ht_params.key_len = fl_mask_range(mask);
1903         mask->filter_ht_params.key_offset += mask->range.start;
1904
1905         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1906 }
1907
1908 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1909 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1910
1911 #define FL_KEY_IS_MASKED(mask, member)                                          \
1912         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1913                    0, FL_KEY_MEMBER_SIZE(member))                               \
1914
1915 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1916         do {                                                                    \
1917                 keys[cnt].key_id = id;                                          \
1918                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1919                 cnt++;                                                          \
1920         } while(0);
1921
1922 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1923         do {                                                                    \
1924                 if (FL_KEY_IS_MASKED(mask, member))                             \
1925                         FL_KEY_SET(keys, cnt, id, member);                      \
1926         } while(0);
1927
1928 static void fl_init_dissector(struct flow_dissector *dissector,
1929                               struct fl_flow_key *mask)
1930 {
1931         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1932         size_t cnt = 0;
1933
1934         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1935                              FLOW_DISSECTOR_KEY_META, meta);
1936         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1937         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1938         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1939                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1940         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1941                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1942         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1943                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1944         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1945                              FLOW_DISSECTOR_KEY_PORTS, tp);
1946         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1947                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1948         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1949                              FLOW_DISSECTOR_KEY_IP, ip);
1950         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1951                              FLOW_DISSECTOR_KEY_TCP, tcp);
1952         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1953                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1954         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1955                              FLOW_DISSECTOR_KEY_ARP, arp);
1956         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1957                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1958         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1959                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1960         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1961                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1962         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1963                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1964         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1965                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1966         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1967                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1968         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1969             FL_KEY_IS_MASKED(mask, enc_ipv6))
1970                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1971                            enc_control);
1972         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1973                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1974         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1975                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1976         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1977                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1978         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1979                              FLOW_DISSECTOR_KEY_CT, ct);
1980         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1981                              FLOW_DISSECTOR_KEY_HASH, hash);
1982         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1983                              FLOW_DISSECTOR_KEY_NUM_OF_VLANS, num_of_vlans);
1984         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1985                              FLOW_DISSECTOR_KEY_PPPOE, pppoe);
1986         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1987                              FLOW_DISSECTOR_KEY_L2TPV3, l2tpv3);
1988
1989         skb_flow_dissector_init(dissector, keys, cnt);
1990 }
1991
1992 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1993                                                struct fl_flow_mask *mask)
1994 {
1995         struct fl_flow_mask *newmask;
1996         int err;
1997
1998         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1999         if (!newmask)
2000                 return ERR_PTR(-ENOMEM);
2001
2002         fl_mask_copy(newmask, mask);
2003
2004         if ((newmask->key.tp_range.tp_min.dst &&
2005              newmask->key.tp_range.tp_max.dst) ||
2006             (newmask->key.tp_range.tp_min.src &&
2007              newmask->key.tp_range.tp_max.src))
2008                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
2009
2010         err = fl_init_mask_hashtable(newmask);
2011         if (err)
2012                 goto errout_free;
2013
2014         fl_init_dissector(&newmask->dissector, &newmask->key);
2015
2016         INIT_LIST_HEAD_RCU(&newmask->filters);
2017
2018         refcount_set(&newmask->refcnt, 1);
2019         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
2020                                       &newmask->ht_node, mask_ht_params);
2021         if (err)
2022                 goto errout_destroy;
2023
2024         spin_lock(&head->masks_lock);
2025         list_add_tail_rcu(&newmask->list, &head->masks);
2026         spin_unlock(&head->masks_lock);
2027
2028         return newmask;
2029
2030 errout_destroy:
2031         rhashtable_destroy(&newmask->ht);
2032 errout_free:
2033         kfree(newmask);
2034
2035         return ERR_PTR(err);
2036 }
2037
2038 static int fl_check_assign_mask(struct cls_fl_head *head,
2039                                 struct cls_fl_filter *fnew,
2040                                 struct cls_fl_filter *fold,
2041                                 struct fl_flow_mask *mask)
2042 {
2043         struct fl_flow_mask *newmask;
2044         int ret = 0;
2045
2046         rcu_read_lock();
2047
2048         /* Insert mask as temporary node to prevent concurrent creation of mask
2049          * with same key. Any concurrent lookups with same key will return
2050          * -EAGAIN because mask's refcnt is zero.
2051          */
2052         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
2053                                                        &mask->ht_node,
2054                                                        mask_ht_params);
2055         if (!fnew->mask) {
2056                 rcu_read_unlock();
2057
2058                 if (fold) {
2059                         ret = -EINVAL;
2060                         goto errout_cleanup;
2061                 }
2062
2063                 newmask = fl_create_new_mask(head, mask);
2064                 if (IS_ERR(newmask)) {
2065                         ret = PTR_ERR(newmask);
2066                         goto errout_cleanup;
2067                 }
2068
2069                 fnew->mask = newmask;
2070                 return 0;
2071         } else if (IS_ERR(fnew->mask)) {
2072                 ret = PTR_ERR(fnew->mask);
2073         } else if (fold && fold->mask != fnew->mask) {
2074                 ret = -EINVAL;
2075         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
2076                 /* Mask was deleted concurrently, try again */
2077                 ret = -EAGAIN;
2078         }
2079         rcu_read_unlock();
2080         return ret;
2081
2082 errout_cleanup:
2083         rhashtable_remove_fast(&head->ht, &mask->ht_node,
2084                                mask_ht_params);
2085         return ret;
2086 }
2087
2088 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
2089                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
2090                         unsigned long base, struct nlattr **tb,
2091                         struct nlattr *est,
2092                         struct fl_flow_tmplt *tmplt,
2093                         u32 flags, u32 fl_flags,
2094                         struct netlink_ext_ack *extack)
2095 {
2096         int err;
2097
2098         err = tcf_exts_validate_ex(net, tp, tb, est, &f->exts, flags,
2099                                    fl_flags, extack);
2100         if (err < 0)
2101                 return err;
2102
2103         if (tb[TCA_FLOWER_CLASSID]) {
2104                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2105                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2106                         rtnl_lock();
2107                 tcf_bind_filter(tp, &f->res, base);
2108                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2109                         rtnl_unlock();
2110         }
2111
2112         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
2113         if (err)
2114                 return err;
2115
2116         fl_mask_update_range(mask);
2117         fl_set_masked_key(&f->mkey, &f->key, mask);
2118
2119         if (!fl_mask_fits_tmplt(tmplt, mask)) {
2120                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2121                 return -EINVAL;
2122         }
2123
2124         return 0;
2125 }
2126
2127 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2128                                struct cls_fl_filter *fold,
2129                                bool *in_ht)
2130 {
2131         struct fl_flow_mask *mask = fnew->mask;
2132         int err;
2133
2134         err = rhashtable_lookup_insert_fast(&mask->ht,
2135                                             &fnew->ht_node,
2136                                             mask->filter_ht_params);
2137         if (err) {
2138                 *in_ht = false;
2139                 /* It is okay if filter with same key exists when
2140                  * overwriting.
2141                  */
2142                 return fold && err == -EEXIST ? 0 : err;
2143         }
2144
2145         *in_ht = true;
2146         return 0;
2147 }
2148
2149 static int fl_change(struct net *net, struct sk_buff *in_skb,
2150                      struct tcf_proto *tp, unsigned long base,
2151                      u32 handle, struct nlattr **tca,
2152                      void **arg, u32 flags,
2153                      struct netlink_ext_ack *extack)
2154 {
2155         struct cls_fl_head *head = fl_head_dereference(tp);
2156         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2157         struct cls_fl_filter *fold = *arg;
2158         struct cls_fl_filter *fnew;
2159         struct fl_flow_mask *mask;
2160         struct nlattr **tb;
2161         bool in_ht;
2162         int err;
2163
2164         if (!tca[TCA_OPTIONS]) {
2165                 err = -EINVAL;
2166                 goto errout_fold;
2167         }
2168
2169         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2170         if (!mask) {
2171                 err = -ENOBUFS;
2172                 goto errout_fold;
2173         }
2174
2175         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2176         if (!tb) {
2177                 err = -ENOBUFS;
2178                 goto errout_mask_alloc;
2179         }
2180
2181         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2182                                           tca[TCA_OPTIONS], fl_policy, NULL);
2183         if (err < 0)
2184                 goto errout_tb;
2185
2186         if (fold && handle && fold->handle != handle) {
2187                 err = -EINVAL;
2188                 goto errout_tb;
2189         }
2190
2191         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2192         if (!fnew) {
2193                 err = -ENOBUFS;
2194                 goto errout_tb;
2195         }
2196         INIT_LIST_HEAD(&fnew->hw_list);
2197         refcount_set(&fnew->refcnt, 1);
2198
2199         if (tb[TCA_FLOWER_FLAGS]) {
2200                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2201
2202                 if (!tc_flags_valid(fnew->flags)) {
2203                         kfree(fnew);
2204                         err = -EINVAL;
2205                         goto errout_tb;
2206                 }
2207         }
2208
2209         if (!fold) {
2210                 spin_lock(&tp->lock);
2211                 if (!handle) {
2212                         handle = 1;
2213                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2214                                             INT_MAX, GFP_ATOMIC);
2215                 } else {
2216                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2217                                             handle, GFP_ATOMIC);
2218
2219                         /* Filter with specified handle was concurrently
2220                          * inserted after initial check in cls_api. This is not
2221                          * necessarily an error if NLM_F_EXCL is not set in
2222                          * message flags. Returning EAGAIN will cause cls_api to
2223                          * try to update concurrently inserted rule.
2224                          */
2225                         if (err == -ENOSPC)
2226                                 err = -EAGAIN;
2227                 }
2228                 spin_unlock(&tp->lock);
2229
2230                 if (err) {
2231                         kfree(fnew);
2232                         goto errout_tb;
2233                 }
2234         }
2235         fnew->handle = handle;
2236
2237         err = tcf_exts_init_ex(&fnew->exts, net, TCA_FLOWER_ACT, 0, tp, handle,
2238                                !tc_skip_hw(fnew->flags));
2239         if (err < 0)
2240                 goto errout_idr;
2241
2242         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2243                            tp->chain->tmplt_priv, flags, fnew->flags,
2244                            extack);
2245         if (err)
2246                 goto errout_idr;
2247
2248         err = fl_check_assign_mask(head, fnew, fold, mask);
2249         if (err)
2250                 goto errout_idr;
2251
2252         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2253         if (err)
2254                 goto errout_mask;
2255
2256         if (!tc_skip_hw(fnew->flags)) {
2257                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2258                 if (err)
2259                         goto errout_ht;
2260         }
2261
2262         if (!tc_in_hw(fnew->flags))
2263                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2264
2265         spin_lock(&tp->lock);
2266
2267         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2268          * proto again or create new one, if necessary.
2269          */
2270         if (tp->deleting) {
2271                 err = -EAGAIN;
2272                 goto errout_hw;
2273         }
2274
2275         if (fold) {
2276                 /* Fold filter was deleted concurrently. Retry lookup. */
2277                 if (fold->deleted) {
2278                         err = -EAGAIN;
2279                         goto errout_hw;
2280                 }
2281
2282                 fnew->handle = handle;
2283
2284                 if (!in_ht) {
2285                         struct rhashtable_params params =
2286                                 fnew->mask->filter_ht_params;
2287
2288                         err = rhashtable_insert_fast(&fnew->mask->ht,
2289                                                      &fnew->ht_node,
2290                                                      params);
2291                         if (err)
2292                                 goto errout_hw;
2293                         in_ht = true;
2294                 }
2295
2296                 refcount_inc(&fnew->refcnt);
2297                 rhashtable_remove_fast(&fold->mask->ht,
2298                                        &fold->ht_node,
2299                                        fold->mask->filter_ht_params);
2300                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2301                 list_replace_rcu(&fold->list, &fnew->list);
2302                 fold->deleted = true;
2303
2304                 spin_unlock(&tp->lock);
2305
2306                 fl_mask_put(head, fold->mask);
2307                 if (!tc_skip_hw(fold->flags))
2308                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2309                 tcf_unbind_filter(tp, &fold->res);
2310                 /* Caller holds reference to fold, so refcnt is always > 0
2311                  * after this.
2312                  */
2313                 refcount_dec(&fold->refcnt);
2314                 __fl_put(fold);
2315         } else {
2316                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2317
2318                 refcount_inc(&fnew->refcnt);
2319                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2320                 spin_unlock(&tp->lock);
2321         }
2322
2323         *arg = fnew;
2324
2325         kfree(tb);
2326         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2327         return 0;
2328
2329 errout_ht:
2330         spin_lock(&tp->lock);
2331 errout_hw:
2332         fnew->deleted = true;
2333         spin_unlock(&tp->lock);
2334         if (!tc_skip_hw(fnew->flags))
2335                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2336         if (in_ht)
2337                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2338                                        fnew->mask->filter_ht_params);
2339 errout_mask:
2340         fl_mask_put(head, fnew->mask);
2341 errout_idr:
2342         idr_remove(&head->handle_idr, fnew->handle);
2343         __fl_put(fnew);
2344 errout_tb:
2345         kfree(tb);
2346 errout_mask_alloc:
2347         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2348 errout_fold:
2349         if (fold)
2350                 __fl_put(fold);
2351         return err;
2352 }
2353
2354 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2355                      bool rtnl_held, struct netlink_ext_ack *extack)
2356 {
2357         struct cls_fl_head *head = fl_head_dereference(tp);
2358         struct cls_fl_filter *f = arg;
2359         bool last_on_mask;
2360         int err = 0;
2361
2362         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2363         *last = list_empty(&head->masks);
2364         __fl_put(f);
2365
2366         return err;
2367 }
2368
2369 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2370                     bool rtnl_held)
2371 {
2372         struct cls_fl_head *head = fl_head_dereference(tp);
2373         unsigned long id = arg->cookie, tmp;
2374         struct cls_fl_filter *f;
2375
2376         arg->count = arg->skip;
2377
2378         rcu_read_lock();
2379         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2380                 /* don't return filters that are being deleted */
2381                 if (!refcount_inc_not_zero(&f->refcnt))
2382                         continue;
2383                 rcu_read_unlock();
2384
2385                 if (arg->fn(tp, f, arg) < 0) {
2386                         __fl_put(f);
2387                         arg->stop = 1;
2388                         rcu_read_lock();
2389                         break;
2390                 }
2391                 __fl_put(f);
2392                 arg->count++;
2393                 rcu_read_lock();
2394         }
2395         rcu_read_unlock();
2396         arg->cookie = id;
2397 }
2398
2399 static struct cls_fl_filter *
2400 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2401 {
2402         struct cls_fl_head *head = fl_head_dereference(tp);
2403
2404         spin_lock(&tp->lock);
2405         if (list_empty(&head->hw_filters)) {
2406                 spin_unlock(&tp->lock);
2407                 return NULL;
2408         }
2409
2410         if (!f)
2411                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2412                                hw_list);
2413         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2414                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2415                         spin_unlock(&tp->lock);
2416                         return f;
2417                 }
2418         }
2419
2420         spin_unlock(&tp->lock);
2421         return NULL;
2422 }
2423
2424 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2425                         void *cb_priv, struct netlink_ext_ack *extack)
2426 {
2427         struct tcf_block *block = tp->chain->block;
2428         struct flow_cls_offload cls_flower = {};
2429         struct cls_fl_filter *f = NULL;
2430         int err;
2431
2432         /* hw_filters list can only be changed by hw offload functions after
2433          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2434          * iterating it.
2435          */
2436         ASSERT_RTNL();
2437
2438         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2439                 cls_flower.rule =
2440                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2441                 if (!cls_flower.rule) {
2442                         __fl_put(f);
2443                         return -ENOMEM;
2444                 }
2445
2446                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2447                                            extack);
2448                 cls_flower.command = add ?
2449                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2450                 cls_flower.cookie = (unsigned long)f;
2451                 cls_flower.rule->match.dissector = &f->mask->dissector;
2452                 cls_flower.rule->match.mask = &f->mask->key;
2453                 cls_flower.rule->match.key = &f->mkey;
2454
2455                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
2456                                               cls_flower.common.extack);
2457                 if (err) {
2458                         kfree(cls_flower.rule);
2459                         if (tc_skip_sw(f->flags)) {
2460                                 __fl_put(f);
2461                                 return err;
2462                         }
2463                         goto next_flow;
2464                 }
2465
2466                 cls_flower.classid = f->res.classid;
2467
2468                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2469                                             TC_SETUP_CLSFLOWER, &cls_flower,
2470                                             cb_priv, &f->flags,
2471                                             &f->in_hw_count);
2472                 tc_cleanup_offload_action(&cls_flower.rule->action);
2473                 kfree(cls_flower.rule);
2474
2475                 if (err) {
2476                         __fl_put(f);
2477                         return err;
2478                 }
2479 next_flow:
2480                 __fl_put(f);
2481         }
2482
2483         return 0;
2484 }
2485
2486 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2487 {
2488         struct flow_cls_offload *cls_flower = type_data;
2489         struct cls_fl_filter *f =
2490                 (struct cls_fl_filter *) cls_flower->cookie;
2491         struct cls_fl_head *head = fl_head_dereference(tp);
2492
2493         spin_lock(&tp->lock);
2494         list_add(&f->hw_list, &head->hw_filters);
2495         spin_unlock(&tp->lock);
2496 }
2497
2498 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2499 {
2500         struct flow_cls_offload *cls_flower = type_data;
2501         struct cls_fl_filter *f =
2502                 (struct cls_fl_filter *) cls_flower->cookie;
2503
2504         spin_lock(&tp->lock);
2505         if (!list_empty(&f->hw_list))
2506                 list_del_init(&f->hw_list);
2507         spin_unlock(&tp->lock);
2508 }
2509
2510 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2511                               struct fl_flow_tmplt *tmplt)
2512 {
2513         struct flow_cls_offload cls_flower = {};
2514         struct tcf_block *block = chain->block;
2515
2516         cls_flower.rule = flow_rule_alloc(0);
2517         if (!cls_flower.rule)
2518                 return -ENOMEM;
2519
2520         cls_flower.common.chain_index = chain->index;
2521         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2522         cls_flower.cookie = (unsigned long) tmplt;
2523         cls_flower.rule->match.dissector = &tmplt->dissector;
2524         cls_flower.rule->match.mask = &tmplt->mask;
2525         cls_flower.rule->match.key = &tmplt->dummy_key;
2526
2527         /* We don't care if driver (any of them) fails to handle this
2528          * call. It serves just as a hint for it.
2529          */
2530         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2531         kfree(cls_flower.rule);
2532
2533         return 0;
2534 }
2535
2536 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2537                                 struct fl_flow_tmplt *tmplt)
2538 {
2539         struct flow_cls_offload cls_flower = {};
2540         struct tcf_block *block = chain->block;
2541
2542         cls_flower.common.chain_index = chain->index;
2543         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2544         cls_flower.cookie = (unsigned long) tmplt;
2545
2546         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2547 }
2548
2549 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2550                              struct nlattr **tca,
2551                              struct netlink_ext_ack *extack)
2552 {
2553         struct fl_flow_tmplt *tmplt;
2554         struct nlattr **tb;
2555         int err;
2556
2557         if (!tca[TCA_OPTIONS])
2558                 return ERR_PTR(-EINVAL);
2559
2560         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2561         if (!tb)
2562                 return ERR_PTR(-ENOBUFS);
2563         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2564                                           tca[TCA_OPTIONS], fl_policy, NULL);
2565         if (err)
2566                 goto errout_tb;
2567
2568         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2569         if (!tmplt) {
2570                 err = -ENOMEM;
2571                 goto errout_tb;
2572         }
2573         tmplt->chain = chain;
2574         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2575         if (err)
2576                 goto errout_tmplt;
2577
2578         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2579
2580         err = fl_hw_create_tmplt(chain, tmplt);
2581         if (err)
2582                 goto errout_tmplt;
2583
2584         kfree(tb);
2585         return tmplt;
2586
2587 errout_tmplt:
2588         kfree(tmplt);
2589 errout_tb:
2590         kfree(tb);
2591         return ERR_PTR(err);
2592 }
2593
2594 static void fl_tmplt_destroy(void *tmplt_priv)
2595 {
2596         struct fl_flow_tmplt *tmplt = tmplt_priv;
2597
2598         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2599         kfree(tmplt);
2600 }
2601
2602 static int fl_dump_key_val(struct sk_buff *skb,
2603                            void *val, int val_type,
2604                            void *mask, int mask_type, int len)
2605 {
2606         int err;
2607
2608         if (!memchr_inv(mask, 0, len))
2609                 return 0;
2610         err = nla_put(skb, val_type, len, val);
2611         if (err)
2612                 return err;
2613         if (mask_type != TCA_FLOWER_UNSPEC) {
2614                 err = nla_put(skb, mask_type, len, mask);
2615                 if (err)
2616                         return err;
2617         }
2618         return 0;
2619 }
2620
2621 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2622                                   struct fl_flow_key *mask)
2623 {
2624         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2625                             TCA_FLOWER_KEY_PORT_DST_MIN,
2626                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2627                             sizeof(key->tp_range.tp_min.dst)) ||
2628             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2629                             TCA_FLOWER_KEY_PORT_DST_MAX,
2630                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2631                             sizeof(key->tp_range.tp_max.dst)) ||
2632             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2633                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2634                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2635                             sizeof(key->tp_range.tp_min.src)) ||
2636             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2637                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2638                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2639                             sizeof(key->tp_range.tp_max.src)))
2640                 return -1;
2641
2642         return 0;
2643 }
2644
2645 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2646                                     struct flow_dissector_key_mpls *mpls_key,
2647                                     struct flow_dissector_key_mpls *mpls_mask,
2648                                     u8 lse_index)
2649 {
2650         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2651         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2652         int err;
2653
2654         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2655                          lse_index + 1);
2656         if (err)
2657                 return err;
2658
2659         if (lse_mask->mpls_ttl) {
2660                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2661                                  lse_key->mpls_ttl);
2662                 if (err)
2663                         return err;
2664         }
2665         if (lse_mask->mpls_bos) {
2666                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2667                                  lse_key->mpls_bos);
2668                 if (err)
2669                         return err;
2670         }
2671         if (lse_mask->mpls_tc) {
2672                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2673                                  lse_key->mpls_tc);
2674                 if (err)
2675                         return err;
2676         }
2677         if (lse_mask->mpls_label) {
2678                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2679                                   lse_key->mpls_label);
2680                 if (err)
2681                         return err;
2682         }
2683
2684         return 0;
2685 }
2686
2687 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2688                                  struct flow_dissector_key_mpls *mpls_key,
2689                                  struct flow_dissector_key_mpls *mpls_mask)
2690 {
2691         struct nlattr *opts;
2692         struct nlattr *lse;
2693         u8 lse_index;
2694         int err;
2695
2696         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2697         if (!opts)
2698                 return -EMSGSIZE;
2699
2700         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2701                 if (!(mpls_mask->used_lses & 1 << lse_index))
2702                         continue;
2703
2704                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2705                 if (!lse) {
2706                         err = -EMSGSIZE;
2707                         goto err_opts;
2708                 }
2709
2710                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2711                                                lse_index);
2712                 if (err)
2713                         goto err_opts_lse;
2714                 nla_nest_end(skb, lse);
2715         }
2716         nla_nest_end(skb, opts);
2717
2718         return 0;
2719
2720 err_opts_lse:
2721         nla_nest_cancel(skb, lse);
2722 err_opts:
2723         nla_nest_cancel(skb, opts);
2724
2725         return err;
2726 }
2727
2728 static int fl_dump_key_mpls(struct sk_buff *skb,
2729                             struct flow_dissector_key_mpls *mpls_key,
2730                             struct flow_dissector_key_mpls *mpls_mask)
2731 {
2732         struct flow_dissector_mpls_lse *lse_mask;
2733         struct flow_dissector_mpls_lse *lse_key;
2734         int err;
2735
2736         if (!mpls_mask->used_lses)
2737                 return 0;
2738
2739         lse_mask = &mpls_mask->ls[0];
2740         lse_key = &mpls_key->ls[0];
2741
2742         /* For backward compatibility, don't use the MPLS nested attributes if
2743          * the rule can be expressed using the old attributes.
2744          */
2745         if (mpls_mask->used_lses & ~1 ||
2746             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2747              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2748                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2749
2750         if (lse_mask->mpls_ttl) {
2751                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2752                                  lse_key->mpls_ttl);
2753                 if (err)
2754                         return err;
2755         }
2756         if (lse_mask->mpls_tc) {
2757                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2758                                  lse_key->mpls_tc);
2759                 if (err)
2760                         return err;
2761         }
2762         if (lse_mask->mpls_label) {
2763                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2764                                   lse_key->mpls_label);
2765                 if (err)
2766                         return err;
2767         }
2768         if (lse_mask->mpls_bos) {
2769                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2770                                  lse_key->mpls_bos);
2771                 if (err)
2772                         return err;
2773         }
2774         return 0;
2775 }
2776
2777 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2778                           struct flow_dissector_key_ip *key,
2779                           struct flow_dissector_key_ip *mask)
2780 {
2781         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2782         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2783         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2784         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2785
2786         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2787             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2788                 return -1;
2789
2790         return 0;
2791 }
2792
2793 static int fl_dump_key_vlan(struct sk_buff *skb,
2794                             int vlan_id_key, int vlan_prio_key,
2795                             struct flow_dissector_key_vlan *vlan_key,
2796                             struct flow_dissector_key_vlan *vlan_mask)
2797 {
2798         int err;
2799
2800         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2801                 return 0;
2802         if (vlan_mask->vlan_id) {
2803                 err = nla_put_u16(skb, vlan_id_key,
2804                                   vlan_key->vlan_id);
2805                 if (err)
2806                         return err;
2807         }
2808         if (vlan_mask->vlan_priority) {
2809                 err = nla_put_u8(skb, vlan_prio_key,
2810                                  vlan_key->vlan_priority);
2811                 if (err)
2812                         return err;
2813         }
2814         return 0;
2815 }
2816
2817 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2818                             u32 *flower_key, u32 *flower_mask,
2819                             u32 flower_flag_bit, u32 dissector_flag_bit)
2820 {
2821         if (dissector_mask & dissector_flag_bit) {
2822                 *flower_mask |= flower_flag_bit;
2823                 if (dissector_key & dissector_flag_bit)
2824                         *flower_key |= flower_flag_bit;
2825         }
2826 }
2827
2828 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2829 {
2830         u32 key, mask;
2831         __be32 _key, _mask;
2832         int err;
2833
2834         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2835                 return 0;
2836
2837         key = 0;
2838         mask = 0;
2839
2840         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2841                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2842         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2843                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2844                         FLOW_DIS_FIRST_FRAG);
2845
2846         _key = cpu_to_be32(key);
2847         _mask = cpu_to_be32(mask);
2848
2849         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2850         if (err)
2851                 return err;
2852
2853         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2854 }
2855
2856 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2857                                   struct flow_dissector_key_enc_opts *enc_opts)
2858 {
2859         struct geneve_opt *opt;
2860         struct nlattr *nest;
2861         int opt_off = 0;
2862
2863         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2864         if (!nest)
2865                 goto nla_put_failure;
2866
2867         while (enc_opts->len > opt_off) {
2868                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2869
2870                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2871                                  opt->opt_class))
2872                         goto nla_put_failure;
2873                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2874                                opt->type))
2875                         goto nla_put_failure;
2876                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2877                             opt->length * 4, opt->opt_data))
2878                         goto nla_put_failure;
2879
2880                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2881         }
2882         nla_nest_end(skb, nest);
2883         return 0;
2884
2885 nla_put_failure:
2886         nla_nest_cancel(skb, nest);
2887         return -EMSGSIZE;
2888 }
2889
2890 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2891                                  struct flow_dissector_key_enc_opts *enc_opts)
2892 {
2893         struct vxlan_metadata *md;
2894         struct nlattr *nest;
2895
2896         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2897         if (!nest)
2898                 goto nla_put_failure;
2899
2900         md = (struct vxlan_metadata *)&enc_opts->data[0];
2901         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2902                 goto nla_put_failure;
2903
2904         nla_nest_end(skb, nest);
2905         return 0;
2906
2907 nla_put_failure:
2908         nla_nest_cancel(skb, nest);
2909         return -EMSGSIZE;
2910 }
2911
2912 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2913                                   struct flow_dissector_key_enc_opts *enc_opts)
2914 {
2915         struct erspan_metadata *md;
2916         struct nlattr *nest;
2917
2918         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2919         if (!nest)
2920                 goto nla_put_failure;
2921
2922         md = (struct erspan_metadata *)&enc_opts->data[0];
2923         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2924                 goto nla_put_failure;
2925
2926         if (md->version == 1 &&
2927             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2928                 goto nla_put_failure;
2929
2930         if (md->version == 2 &&
2931             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2932                         md->u.md2.dir) ||
2933              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2934                         get_hwid(&md->u.md2))))
2935                 goto nla_put_failure;
2936
2937         nla_nest_end(skb, nest);
2938         return 0;
2939
2940 nla_put_failure:
2941         nla_nest_cancel(skb, nest);
2942         return -EMSGSIZE;
2943 }
2944
2945 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
2946                                struct flow_dissector_key_enc_opts *enc_opts)
2947
2948 {
2949         struct gtp_pdu_session_info *session_info;
2950         struct nlattr *nest;
2951
2952         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
2953         if (!nest)
2954                 goto nla_put_failure;
2955
2956         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
2957
2958         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
2959                        session_info->pdu_type))
2960                 goto nla_put_failure;
2961
2962         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
2963                 goto nla_put_failure;
2964
2965         nla_nest_end(skb, nest);
2966         return 0;
2967
2968 nla_put_failure:
2969         nla_nest_cancel(skb, nest);
2970         return -EMSGSIZE;
2971 }
2972
2973 static int fl_dump_key_ct(struct sk_buff *skb,
2974                           struct flow_dissector_key_ct *key,
2975                           struct flow_dissector_key_ct *mask)
2976 {
2977         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2978             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2979                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2980                             sizeof(key->ct_state)))
2981                 goto nla_put_failure;
2982
2983         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2984             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2985                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2986                             sizeof(key->ct_zone)))
2987                 goto nla_put_failure;
2988
2989         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2990             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2991                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2992                             sizeof(key->ct_mark)))
2993                 goto nla_put_failure;
2994
2995         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2996             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2997                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2998                             sizeof(key->ct_labels)))
2999                 goto nla_put_failure;
3000
3001         return 0;
3002
3003 nla_put_failure:
3004         return -EMSGSIZE;
3005 }
3006
3007 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
3008                                struct flow_dissector_key_enc_opts *enc_opts)
3009 {
3010         struct nlattr *nest;
3011         int err;
3012
3013         if (!enc_opts->len)
3014                 return 0;
3015
3016         nest = nla_nest_start_noflag(skb, enc_opt_type);
3017         if (!nest)
3018                 goto nla_put_failure;
3019
3020         switch (enc_opts->dst_opt_type) {
3021         case TUNNEL_GENEVE_OPT:
3022                 err = fl_dump_key_geneve_opt(skb, enc_opts);
3023                 if (err)
3024                         goto nla_put_failure;
3025                 break;
3026         case TUNNEL_VXLAN_OPT:
3027                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
3028                 if (err)
3029                         goto nla_put_failure;
3030                 break;
3031         case TUNNEL_ERSPAN_OPT:
3032                 err = fl_dump_key_erspan_opt(skb, enc_opts);
3033                 if (err)
3034                         goto nla_put_failure;
3035                 break;
3036         case TUNNEL_GTP_OPT:
3037                 err = fl_dump_key_gtp_opt(skb, enc_opts);
3038                 if (err)
3039                         goto nla_put_failure;
3040                 break;
3041         default:
3042                 goto nla_put_failure;
3043         }
3044         nla_nest_end(skb, nest);
3045         return 0;
3046
3047 nla_put_failure:
3048         nla_nest_cancel(skb, nest);
3049         return -EMSGSIZE;
3050 }
3051
3052 static int fl_dump_key_enc_opt(struct sk_buff *skb,
3053                                struct flow_dissector_key_enc_opts *key_opts,
3054                                struct flow_dissector_key_enc_opts *msk_opts)
3055 {
3056         int err;
3057
3058         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
3059         if (err)
3060                 return err;
3061
3062         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
3063 }
3064
3065 static int fl_dump_key(struct sk_buff *skb, struct net *net,
3066                        struct fl_flow_key *key, struct fl_flow_key *mask)
3067 {
3068         if (mask->meta.ingress_ifindex) {
3069                 struct net_device *dev;
3070
3071                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
3072                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
3073                         goto nla_put_failure;
3074         }
3075
3076         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
3077                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
3078                             sizeof(key->eth.dst)) ||
3079             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
3080                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
3081                             sizeof(key->eth.src)) ||
3082             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
3083                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
3084                             sizeof(key->basic.n_proto)))
3085                 goto nla_put_failure;
3086
3087         if (mask->num_of_vlans.num_of_vlans) {
3088                 if (nla_put_u8(skb, TCA_FLOWER_KEY_NUM_OF_VLANS, key->num_of_vlans.num_of_vlans))
3089                         goto nla_put_failure;
3090         }
3091
3092         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
3093                 goto nla_put_failure;
3094
3095         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
3096                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
3097                 goto nla_put_failure;
3098
3099         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
3100                              TCA_FLOWER_KEY_CVLAN_PRIO,
3101                              &key->cvlan, &mask->cvlan) ||
3102             (mask->cvlan.vlan_tpid &&
3103              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3104                           key->cvlan.vlan_tpid)))
3105                 goto nla_put_failure;
3106
3107         if (mask->basic.n_proto) {
3108                 if (mask->cvlan.vlan_eth_type) {
3109                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3110                                          key->basic.n_proto))
3111                                 goto nla_put_failure;
3112                 } else if (mask->vlan.vlan_eth_type) {
3113                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3114                                          key->vlan.vlan_eth_type))
3115                                 goto nla_put_failure;
3116                 }
3117         }
3118
3119         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3120              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3121             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3122                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3123                             sizeof(key->basic.ip_proto)) ||
3124             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3125                 goto nla_put_failure;
3126
3127         if (mask->pppoe.session_id) {
3128                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPPOE_SID,
3129                                  key->pppoe.session_id))
3130                         goto nla_put_failure;
3131         }
3132         if (mask->basic.n_proto && mask->pppoe.ppp_proto) {
3133                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPP_PROTO,
3134                                  key->pppoe.ppp_proto))
3135                         goto nla_put_failure;
3136         }
3137
3138         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3139             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3140                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3141                              sizeof(key->ipv4.src)) ||
3142              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3143                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3144                              sizeof(key->ipv4.dst))))
3145                 goto nla_put_failure;
3146         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3147                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3148                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3149                                   sizeof(key->ipv6.src)) ||
3150                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3151                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3152                                   sizeof(key->ipv6.dst))))
3153                 goto nla_put_failure;
3154
3155         if (key->basic.ip_proto == IPPROTO_TCP &&
3156             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3157                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3158                              sizeof(key->tp.src)) ||
3159              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3160                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3161                              sizeof(key->tp.dst)) ||
3162              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3163                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3164                              sizeof(key->tcp.flags))))
3165                 goto nla_put_failure;
3166         else if (key->basic.ip_proto == IPPROTO_UDP &&
3167                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3168                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3169                                   sizeof(key->tp.src)) ||
3170                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3171                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3172                                   sizeof(key->tp.dst))))
3173                 goto nla_put_failure;
3174         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3175                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3176                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3177                                   sizeof(key->tp.src)) ||
3178                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3179                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3180                                   sizeof(key->tp.dst))))
3181                 goto nla_put_failure;
3182         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3183                  key->basic.ip_proto == IPPROTO_ICMP &&
3184                  (fl_dump_key_val(skb, &key->icmp.type,
3185                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3186                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3187                                   sizeof(key->icmp.type)) ||
3188                   fl_dump_key_val(skb, &key->icmp.code,
3189                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3190                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3191                                   sizeof(key->icmp.code))))
3192                 goto nla_put_failure;
3193         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3194                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3195                  (fl_dump_key_val(skb, &key->icmp.type,
3196                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3197                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3198                                   sizeof(key->icmp.type)) ||
3199                   fl_dump_key_val(skb, &key->icmp.code,
3200                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3201                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3202                                   sizeof(key->icmp.code))))
3203                 goto nla_put_failure;
3204         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3205                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3206                  (fl_dump_key_val(skb, &key->arp.sip,
3207                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3208                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3209                                   sizeof(key->arp.sip)) ||
3210                   fl_dump_key_val(skb, &key->arp.tip,
3211                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3212                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3213                                   sizeof(key->arp.tip)) ||
3214                   fl_dump_key_val(skb, &key->arp.op,
3215                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3216                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3217                                   sizeof(key->arp.op)) ||
3218                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3219                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3220                                   sizeof(key->arp.sha)) ||
3221                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3222                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3223                                   sizeof(key->arp.tha))))
3224                 goto nla_put_failure;
3225         else if (key->basic.ip_proto == IPPROTO_L2TP &&
3226                  fl_dump_key_val(skb, &key->l2tpv3.session_id,
3227                                  TCA_FLOWER_KEY_L2TPV3_SID,
3228                                  &mask->l2tpv3.session_id,
3229                                  TCA_FLOWER_UNSPEC,
3230                                  sizeof(key->l2tpv3.session_id)))
3231                 goto nla_put_failure;
3232
3233         if ((key->basic.ip_proto == IPPROTO_TCP ||
3234              key->basic.ip_proto == IPPROTO_UDP ||
3235              key->basic.ip_proto == IPPROTO_SCTP) &&
3236              fl_dump_key_port_range(skb, key, mask))
3237                 goto nla_put_failure;
3238
3239         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3240             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3241                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3242                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3243                             sizeof(key->enc_ipv4.src)) ||
3244              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3245                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3246                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3247                              sizeof(key->enc_ipv4.dst))))
3248                 goto nla_put_failure;
3249         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3250                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3251                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3252                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3253                             sizeof(key->enc_ipv6.src)) ||
3254                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3255                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3256                                  &mask->enc_ipv6.dst,
3257                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3258                             sizeof(key->enc_ipv6.dst))))
3259                 goto nla_put_failure;
3260
3261         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3262                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3263                             sizeof(key->enc_key_id)) ||
3264             fl_dump_key_val(skb, &key->enc_tp.src,
3265                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3266                             &mask->enc_tp.src,
3267                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3268                             sizeof(key->enc_tp.src)) ||
3269             fl_dump_key_val(skb, &key->enc_tp.dst,
3270                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3271                             &mask->enc_tp.dst,
3272                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3273                             sizeof(key->enc_tp.dst)) ||
3274             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3275             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3276                 goto nla_put_failure;
3277
3278         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3279                 goto nla_put_failure;
3280
3281         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3282                 goto nla_put_failure;
3283
3284         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3285                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3286                              sizeof(key->hash.hash)))
3287                 goto nla_put_failure;
3288
3289         return 0;
3290
3291 nla_put_failure:
3292         return -EMSGSIZE;
3293 }
3294
3295 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3296                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3297 {
3298         struct cls_fl_filter *f = fh;
3299         struct nlattr *nest;
3300         struct fl_flow_key *key, *mask;
3301         bool skip_hw;
3302
3303         if (!f)
3304                 return skb->len;
3305
3306         t->tcm_handle = f->handle;
3307
3308         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3309         if (!nest)
3310                 goto nla_put_failure;
3311
3312         spin_lock(&tp->lock);
3313
3314         if (f->res.classid &&
3315             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3316                 goto nla_put_failure_locked;
3317
3318         key = &f->key;
3319         mask = &f->mask->key;
3320         skip_hw = tc_skip_hw(f->flags);
3321
3322         if (fl_dump_key(skb, net, key, mask))
3323                 goto nla_put_failure_locked;
3324
3325         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3326                 goto nla_put_failure_locked;
3327
3328         spin_unlock(&tp->lock);
3329
3330         if (!skip_hw)
3331                 fl_hw_update_stats(tp, f, rtnl_held);
3332
3333         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3334                 goto nla_put_failure;
3335
3336         if (tcf_exts_dump(skb, &f->exts))
3337                 goto nla_put_failure;
3338
3339         nla_nest_end(skb, nest);
3340
3341         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3342                 goto nla_put_failure;
3343
3344         return skb->len;
3345
3346 nla_put_failure_locked:
3347         spin_unlock(&tp->lock);
3348 nla_put_failure:
3349         nla_nest_cancel(skb, nest);
3350         return -1;
3351 }
3352
3353 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3354                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3355 {
3356         struct cls_fl_filter *f = fh;
3357         struct nlattr *nest;
3358         bool skip_hw;
3359
3360         if (!f)
3361                 return skb->len;
3362
3363         t->tcm_handle = f->handle;
3364
3365         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3366         if (!nest)
3367                 goto nla_put_failure;
3368
3369         spin_lock(&tp->lock);
3370
3371         skip_hw = tc_skip_hw(f->flags);
3372
3373         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3374                 goto nla_put_failure_locked;
3375
3376         spin_unlock(&tp->lock);
3377
3378         if (!skip_hw)
3379                 fl_hw_update_stats(tp, f, rtnl_held);
3380
3381         if (tcf_exts_terse_dump(skb, &f->exts))
3382                 goto nla_put_failure;
3383
3384         nla_nest_end(skb, nest);
3385
3386         return skb->len;
3387
3388 nla_put_failure_locked:
3389         spin_unlock(&tp->lock);
3390 nla_put_failure:
3391         nla_nest_cancel(skb, nest);
3392         return -1;
3393 }
3394
3395 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3396 {
3397         struct fl_flow_tmplt *tmplt = tmplt_priv;
3398         struct fl_flow_key *key, *mask;
3399         struct nlattr *nest;
3400
3401         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3402         if (!nest)
3403                 goto nla_put_failure;
3404
3405         key = &tmplt->dummy_key;
3406         mask = &tmplt->mask;
3407
3408         if (fl_dump_key(skb, net, key, mask))
3409                 goto nla_put_failure;
3410
3411         nla_nest_end(skb, nest);
3412
3413         return skb->len;
3414
3415 nla_put_failure:
3416         nla_nest_cancel(skb, nest);
3417         return -EMSGSIZE;
3418 }
3419
3420 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3421                           unsigned long base)
3422 {
3423         struct cls_fl_filter *f = fh;
3424
3425         tc_cls_bind_class(classid, cl, q, &f->res, base);
3426 }
3427
3428 static bool fl_delete_empty(struct tcf_proto *tp)
3429 {
3430         struct cls_fl_head *head = fl_head_dereference(tp);
3431
3432         spin_lock(&tp->lock);
3433         tp->deleting = idr_is_empty(&head->handle_idr);
3434         spin_unlock(&tp->lock);
3435
3436         return tp->deleting;
3437 }
3438
3439 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3440         .kind           = "flower",
3441         .classify       = fl_classify,
3442         .init           = fl_init,
3443         .destroy        = fl_destroy,
3444         .get            = fl_get,
3445         .put            = fl_put,
3446         .change         = fl_change,
3447         .delete         = fl_delete,
3448         .delete_empty   = fl_delete_empty,
3449         .walk           = fl_walk,
3450         .reoffload      = fl_reoffload,
3451         .hw_add         = fl_hw_add,
3452         .hw_del         = fl_hw_del,
3453         .dump           = fl_dump,
3454         .terse_dump     = fl_terse_dump,
3455         .bind_class     = fl_bind_class,
3456         .tmplt_create   = fl_tmplt_create,
3457         .tmplt_destroy  = fl_tmplt_destroy,
3458         .tmplt_dump     = fl_tmplt_dump,
3459         .get_exts       = fl_get_exts,
3460         .owner          = THIS_MODULE,
3461         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3462 };
3463
3464 static int __init cls_fl_init(void)
3465 {
3466         return register_tcf_proto_ops(&cls_fl_ops);
3467 }
3468
3469 static void __exit cls_fl_exit(void)
3470 {
3471         unregister_tcf_proto_ops(&cls_fl_ops);
3472 }
3473
3474 module_init(cls_fl_init);
3475 module_exit(cls_fl_exit);
3476
3477 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3478 MODULE_DESCRIPTION("Flower classifier");
3479 MODULE_LICENSE("GPL v2");