net/sched/cls_flower.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/cls_flower.c               Flower classifier
   4  *
   5  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/rhashtable.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/refcount.h>
  14
  15 #include <linux/if_ether.h>
  16 #include <linux/in6.h>
  17 #include <linux/ip.h>
  18 #include <linux/mpls.h>
  19
  20 #include <net/sch_generic.h>
  21 #include <net/pkt_cls.h>
  22 #include <net/ip.h>
  23 #include <net/flow_dissector.h>
  24 #include <net/geneve.h>
  25
  26 #include <net/dst.h>
  27 #include <net/dst_metadata.h>
  28
  29 struct fl_flow_key {
  30         int     indev_ifindex;
  31         struct flow_dissector_key_control control;
  32         struct flow_dissector_key_control enc_control;
  33         struct flow_dissector_key_basic basic;
  34         struct flow_dissector_key_eth_addrs eth;
  35         struct flow_dissector_key_vlan vlan;
  36         struct flow_dissector_key_vlan cvlan;
  37         union {
  38                 struct flow_dissector_key_ipv4_addrs ipv4;
  39                 struct flow_dissector_key_ipv6_addrs ipv6;
  40         };
  41         struct flow_dissector_key_ports tp;
  42         struct flow_dissector_key_icmp icmp;
  43         struct flow_dissector_key_arp arp;
  44         struct flow_dissector_key_keyid enc_key_id;
  45         union {
  46                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  47                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  48         };
  49         struct flow_dissector_key_ports enc_tp;
  50         struct flow_dissector_key_mpls mpls;
  51         struct flow_dissector_key_tcp tcp;
  52         struct flow_dissector_key_ip ip;
  53         struct flow_dissector_key_ip enc_ip;
  54         struct flow_dissector_key_enc_opts enc_opts;
  55         struct flow_dissector_key_ports tp_min;
  56         struct flow_dissector_key_ports tp_max;
  57 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  58
  59 struct fl_flow_mask_range {
  60         unsigned short int start;
  61         unsigned short int end;
  62 };
  63
  64 struct fl_flow_mask {
  65         struct fl_flow_key key;
  66         struct fl_flow_mask_range range;
  67         u32 flags;
  68         struct rhash_head ht_node;
  69         struct rhashtable ht;
  70         struct rhashtable_params filter_ht_params;
  71         struct flow_dissector dissector;
  72         struct list_head filters;
  73         struct rcu_work rwork;
  74         struct list_head list;
  75         refcount_t refcnt;
  76 };
  77
  78 struct fl_flow_tmplt {
  79         struct fl_flow_key dummy_key;
  80         struct fl_flow_key mask;
  81         struct flow_dissector dissector;
  82         struct tcf_chain *chain;
  83 };
  84
  85 struct cls_fl_head {
  86         struct rhashtable ht;
  87         spinlock_t masks_lock; /* Protect masks list */
  88         struct list_head masks;
  89         struct list_head hw_filters;
  90         struct rcu_work rwork;
  91         struct idr handle_idr;
  92 };
  93
  94 struct cls_fl_filter {
  95         struct fl_flow_mask *mask;
  96         struct rhash_head ht_node;
  97         struct fl_flow_key mkey;
  98         struct tcf_exts exts;
  99         struct tcf_result res;
 100         struct fl_flow_key key;
 101         struct list_head list;
 102         struct list_head hw_list;
 103         u32 handle;
 104         u32 flags;
 105         u32 in_hw_count;
 106         struct rcu_work rwork;
 107         struct net_device *hw_dev;
 108         /* Flower classifier is unlocked, which means that its reference counter
 109          * can be changed concurrently without any kind of external
 110          * synchronization. Use atomic reference counter to be concurrency-safe.
 111          */
 112         refcount_t refcnt;
 113         bool deleted;
 114 };
 115
 116 static const struct rhashtable_params mask_ht_params = {
 117         .key_offset = offsetof(struct fl_flow_mask, key),
 118         .key_len = sizeof(struct fl_flow_key),
 119         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 120         .automatic_shrinking = true,
 121 };
 122
 123 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 124 {
 125         return mask->range.end - mask->range.start;
 126 }
 127
 128 static void fl_mask_update_range(struct fl_flow_mask *mask)
 129 {
 130         const u8 *bytes = (const u8 *) &mask->key;
 131         size_t size = sizeof(mask->key);
 132         size_t i, first = 0, last;
 133
 134         for (i = 0; i < size; i++) {
 135                 if (bytes[i]) {
 136                         first = i;
 137                         break;
 138                 }
 139         }
 140         last = first;
 141         for (i = size - 1; i != first; i--) {
 142                 if (bytes[i]) {
 143                         last = i;
 144                         break;
 145                 }
 146         }
 147         mask->range.start = rounddown(first, sizeof(long));
 148         mask->range.end = roundup(last + 1, sizeof(long));
 149 }
 150
 151 static void *fl_key_get_start(struct fl_flow_key *key,
 152                               const struct fl_flow_mask *mask)
 153 {
 154         return (u8 *) key + mask->range.start;
 155 }
 156
 157 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 158                               struct fl_flow_mask *mask)
 159 {
 160         const long *lkey = fl_key_get_start(key, mask);
 161         const long *lmask = fl_key_get_start(&mask->key, mask);
 162         long *lmkey = fl_key_get_start(mkey, mask);
 163         int i;
 164
 165         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 166                 *lmkey++ = *lkey++ & *lmask++;
 167 }
 168
 169 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 170                                struct fl_flow_mask *mask)
 171 {
 172         const long *lmask = fl_key_get_start(&mask->key, mask);
 173         const long *ltmplt;
 174         int i;
 175
 176         if (!tmplt)
 177                 return true;
 178         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 179         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 180                 if (~*ltmplt++ & *lmask++)
 181                         return false;
 182         }
 183         return true;
 184 }
 185
 186 static void fl_clear_masked_range(struct fl_flow_key *key,
 187                                   struct fl_flow_mask *mask)
 188 {
 189         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 190 }
 191
 192 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 193                                   struct fl_flow_key *key,
 194                                   struct fl_flow_key *mkey)
 195 {
 196         __be16 min_mask, max_mask, min_val, max_val;
 197
 198         min_mask = htons(filter->mask->key.tp_min.dst);
 199         max_mask = htons(filter->mask->key.tp_max.dst);
 200         min_val = htons(filter->key.tp_min.dst);
 201         max_val = htons(filter->key.tp_max.dst);
 202
 203         if (min_mask && max_mask) {
 204                 if (htons(key->tp.dst) < min_val ||
 205                     htons(key->tp.dst) > max_val)
 206                         return false;
 207
 208                 /* skb does not have min and max values */
 209                 mkey->tp_min.dst = filter->mkey.tp_min.dst;
 210                 mkey->tp_max.dst = filter->mkey.tp_max.dst;
 211         }
 212         return true;
 213 }
 214
 215 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 216                                   struct fl_flow_key *key,
 217                                   struct fl_flow_key *mkey)
 218 {
 219         __be16 min_mask, max_mask, min_val, max_val;
 220
 221         min_mask = htons(filter->mask->key.tp_min.src);
 222         max_mask = htons(filter->mask->key.tp_max.src);
 223         min_val = htons(filter->key.tp_min.src);
 224         max_val = htons(filter->key.tp_max.src);
 225
 226         if (min_mask && max_mask) {
 227                 if (htons(key->tp.src) < min_val ||
 228                     htons(key->tp.src) > max_val)
 229                         return false;
 230
 231                 /* skb does not have min and max values */
 232                 mkey->tp_min.src = filter->mkey.tp_min.src;
 233                 mkey->tp_max.src = filter->mkey.tp_max.src;
 234         }
 235         return true;
 236 }
 237
 238 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 239                                          struct fl_flow_key *mkey)
 240 {
 241         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 242                                       mask->filter_ht_params);
 243 }
 244
 245 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 246                                              struct fl_flow_key *mkey,
 247                                              struct fl_flow_key *key)
 248 {
 249         struct cls_fl_filter *filter, *f;
 250
 251         list_for_each_entry_rcu(filter, &mask->filters, list) {
 252                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 253                         continue;
 254
 255                 if (!fl_range_port_src_cmp(filter, key, mkey))
 256                         continue;
 257
 258                 f = __fl_lookup(mask, mkey);
 259                 if (f)
 260                         return f;
 261         }
 262         return NULL;
 263 }
 264
 265 static struct cls_fl_filter *fl_lookup(struct fl_flow_mask *mask,
 266                                        struct fl_flow_key *mkey,
 267                                        struct fl_flow_key *key)
 268 {
 269         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 270                 return fl_lookup_range(mask, mkey, key);
 271
 272         return __fl_lookup(mask, mkey);
 273 }
 274
 275 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 276                        struct tcf_result *res)
 277 {
 278         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 279         struct cls_fl_filter *f;
 280         struct fl_flow_mask *mask;
 281         struct fl_flow_key skb_key;
 282         struct fl_flow_key skb_mkey;
 283
 284         list_for_each_entry_rcu(mask, &head->masks, list) {
 285                 fl_clear_masked_range(&skb_key, mask);
 286
 287                 skb_key.indev_ifindex = skb->skb_iif;
 288                 /* skb_flow_dissect() does not set n_proto in case an unknown
 289                  * protocol, so do it rather here.
 290                  */
 291                 skb_key.basic.n_proto = skb->protocol;
 292                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 293                 skb_flow_dissect(skb, &mask->dissector, &skb_key, 0);
 294
 295                 fl_set_masked_key(&skb_mkey, &skb_key, mask);
 296
 297                 f = fl_lookup(mask, &skb_mkey, &skb_key);
 298                 if (f && !tc_skip_sw(f->flags)) {
 299                         *res = f->res;
 300                         return tcf_exts_exec(skb, &f->exts, res);
 301                 }
 302         }
 303         return -1;
 304 }
 305
 306 static int fl_init(struct tcf_proto *tp)
 307 {
 308         struct cls_fl_head *head;
 309
 310         head = kzalloc(sizeof(*head), GFP_KERNEL);
 311         if (!head)
 312                 return -ENOBUFS;
 313
 314         spin_lock_init(&head->masks_lock);
 315         INIT_LIST_HEAD_RCU(&head->masks);
 316         INIT_LIST_HEAD(&head->hw_filters);
 317         rcu_assign_pointer(tp->root, head);
 318         idr_init(&head->handle_idr);
 319
 320         return rhashtable_init(&head->ht, &mask_ht_params);
 321 }
 322
 323 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 324 {
 325         /* temporary masks don't have their filters list and ht initialized */
 326         if (mask_init_done) {
 327                 WARN_ON(!list_empty(&mask->filters));
 328                 rhashtable_destroy(&mask->ht);
 329         }
 330         kfree(mask);
 331 }
 332
 333 static void fl_mask_free_work(struct work_struct *work)
 334 {
 335         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 336                                                  struct fl_flow_mask, rwork);
 337
 338         fl_mask_free(mask, true);
 339 }
 340
 341 static void fl_uninit_mask_free_work(struct work_struct *work)
 342 {
 343         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 344                                                  struct fl_flow_mask, rwork);
 345
 346         fl_mask_free(mask, false);
 347 }
 348
 349 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 350 {
 351         if (!refcount_dec_and_test(&mask->refcnt))
 352                 return false;
 353
 354         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 355
 356         spin_lock(&head->masks_lock);
 357         list_del_rcu(&mask->list);
 358         spin_unlock(&head->masks_lock);
 359
 360         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 361
 362         return true;
 363 }
 364
 365 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 366 {
 367         /* Flower classifier only changes root pointer during init and destroy.
 368          * Users must obtain reference to tcf_proto instance before calling its
 369          * API, so tp->root pointer is protected from concurrent call to
 370          * fl_destroy() by reference counting.
 371          */
 372         return rcu_dereference_raw(tp->root);
 373 }
 374
 375 static void __fl_destroy_filter(struct cls_fl_filter *f)
 376 {
 377         tcf_exts_destroy(&f->exts);
 378         tcf_exts_put_net(&f->exts);
 379         kfree(f);
 380 }
 381
 382 static void fl_destroy_filter_work(struct work_struct *work)
 383 {
 384         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 385                                         struct cls_fl_filter, rwork);
 386
 387         __fl_destroy_filter(f);
 388 }
 389
 390 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 391                                  bool rtnl_held, struct netlink_ext_ack *extack)
 392 {
 393         struct tc_cls_flower_offload cls_flower = {};
 394         struct tcf_block *block = tp->chain->block;
 395
 396         if (!rtnl_held)
 397                 rtnl_lock();
 398
 399         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 400         cls_flower.command = TC_CLSFLOWER_DESTROY;
 401         cls_flower.cookie = (unsigned long) f;
 402
 403         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
 404         spin_lock(&tp->lock);
 405         list_del_init(&f->hw_list);
 406         tcf_block_offload_dec(block, &f->flags);
 407         spin_unlock(&tp->lock);
 408
 409         if (!rtnl_held)
 410                 rtnl_unlock();
 411 }
 412
 413 static int fl_hw_replace_filter(struct tcf_proto *tp,
 414                                 struct cls_fl_filter *f, bool rtnl_held,
 415                                 struct netlink_ext_ack *extack)
 416 {
 417         struct cls_fl_head *head = fl_head_dereference(tp);
 418         struct tc_cls_flower_offload cls_flower = {};
 419         struct tcf_block *block = tp->chain->block;
 420         bool skip_sw = tc_skip_sw(f->flags);
 421         int err = 0;
 422
 423         if (!rtnl_held)
 424                 rtnl_lock();
 425
 426         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 427         if (!cls_flower.rule) {
 428                 err = -ENOMEM;
 429                 goto errout;
 430         }
 431
 432         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 433         cls_flower.command = TC_CLSFLOWER_REPLACE;
 434         cls_flower.cookie = (unsigned long) f;
 435         cls_flower.rule->match.dissector = &f->mask->dissector;
 436         cls_flower.rule->match.mask = &f->mask->key;
 437         cls_flower.rule->match.key = &f->mkey;
 438         cls_flower.classid = f->res.classid;
 439
 440         err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
 441         if (err) {
 442                 kfree(cls_flower.rule);
 443                 if (skip_sw)
 444                         NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
 445                 else
 446                         err = 0;
 447                 goto errout;
 448         }
 449
 450         err = tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, skip_sw);
 451         kfree(cls_flower.rule);
 452
 453         if (err < 0) {
 454                 fl_hw_destroy_filter(tp, f, true, NULL);
 455                 goto errout;
 456         } else if (err > 0) {
 457                 f->in_hw_count = err;
 458                 err = 0;
 459                 spin_lock(&tp->lock);
 460                 tcf_block_offload_inc(block, &f->flags);
 461                 spin_unlock(&tp->lock);
 462         }
 463
 464         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW)) {
 465                 err = -EINVAL;
 466                 goto errout;
 467         }
 468
 469         spin_lock(&tp->lock);
 470         list_add(&f->hw_list, &head->hw_filters);
 471         spin_unlock(&tp->lock);
 472 errout:
 473         if (!rtnl_held)
 474                 rtnl_unlock();
 475
 476         return err;
 477 }
 478
 479 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 480                                bool rtnl_held)
 481 {
 482         struct tc_cls_flower_offload cls_flower = {};
 483         struct tcf_block *block = tp->chain->block;
 484
 485         if (!rtnl_held)
 486                 rtnl_lock();
 487
 488         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 489         cls_flower.command = TC_CLSFLOWER_STATS;
 490         cls_flower.cookie = (unsigned long) f;
 491         cls_flower.classid = f->res.classid;
 492
 493         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
 494
 495         tcf_exts_stats_update(&f->exts, cls_flower.stats.bytes,
 496                               cls_flower.stats.pkts,
 497                               cls_flower.stats.lastused);
 498
 499         if (!rtnl_held)
 500                 rtnl_unlock();
 501 }
 502
 503 static void __fl_put(struct cls_fl_filter *f)
 504 {
 505         if (!refcount_dec_and_test(&f->refcnt))
 506                 return;
 507
 508         if (tcf_exts_get_net(&f->exts))
 509                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 510         else
 511                 __fl_destroy_filter(f);
 512 }
 513
 514 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 515 {
 516         struct cls_fl_filter *f;
 517
 518         rcu_read_lock();
 519         f = idr_find(&head->handle_idr, handle);
 520         if (f && !refcount_inc_not_zero(&f->refcnt))
 521                 f = NULL;
 522         rcu_read_unlock();
 523
 524         return f;
 525 }
 526
 527 static struct cls_fl_filter *fl_get_next_filter(struct tcf_proto *tp,
 528                                                 unsigned long *handle)
 529 {
 530         struct cls_fl_head *head = fl_head_dereference(tp);
 531         struct cls_fl_filter *f;
 532
 533         rcu_read_lock();
 534         while ((f = idr_get_next_ul(&head->handle_idr, handle))) {
 535                 /* don't return filters that are being deleted */
 536                 if (refcount_inc_not_zero(&f->refcnt))
 537                         break;
 538                 ++(*handle);
 539         }
 540         rcu_read_unlock();
 541
 542         return f;
 543 }
 544
 545 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 546                        bool *last, bool rtnl_held,
 547                        struct netlink_ext_ack *extack)
 548 {
 549         struct cls_fl_head *head = fl_head_dereference(tp);
 550
 551         *last = false;
 552
 553         spin_lock(&tp->lock);
 554         if (f->deleted) {
 555                 spin_unlock(&tp->lock);
 556                 return -ENOENT;
 557         }
 558
 559         f->deleted = true;
 560         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 561                                f->mask->filter_ht_params);
 562         idr_remove(&head->handle_idr, f->handle);
 563         list_del_rcu(&f->list);
 564         spin_unlock(&tp->lock);
 565
 566         *last = fl_mask_put(head, f->mask);
 567         if (!tc_skip_hw(f->flags))
 568                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 569         tcf_unbind_filter(tp, &f->res);
 570         __fl_put(f);
 571
 572         return 0;
 573 }
 574
 575 static void fl_destroy_sleepable(struct work_struct *work)
 576 {
 577         struct cls_fl_head *head = container_of(to_rcu_work(work),
 578                                                 struct cls_fl_head,
 579                                                 rwork);
 580
 581         rhashtable_destroy(&head->ht);
 582         kfree(head);
 583         module_put(THIS_MODULE);
 584 }
 585
 586 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 587                        struct netlink_ext_ack *extack)
 588 {
 589         struct cls_fl_head *head = fl_head_dereference(tp);
 590         struct fl_flow_mask *mask, *next_mask;
 591         struct cls_fl_filter *f, *next;
 592         bool last;
 593
 594         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 595                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 596                         __fl_delete(tp, f, &last, rtnl_held, extack);
 597                         if (last)
 598                                 break;
 599                 }
 600         }
 601         idr_destroy(&head->handle_idr);
 602
 603         __module_get(THIS_MODULE);
 604         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 605 }
 606
 607 static void fl_put(struct tcf_proto *tp, void *arg)
 608 {
 609         struct cls_fl_filter *f = arg;
 610
 611         __fl_put(f);
 612 }
 613
 614 static void *fl_get(struct tcf_proto *tp, u32 handle)
 615 {
 616         struct cls_fl_head *head = fl_head_dereference(tp);
 617
 618         return __fl_get(head, handle);
 619 }
 620
 621 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 622         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 623         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 624         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 625                                             .len = IFNAMSIZ },
 626         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 627         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 628         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 629         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 630         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 631         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 632         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 633         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 634         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 635         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 636         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 637         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 638         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 639         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 640         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 641         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 642         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 643         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 644         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 645         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 646         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 647         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 648         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 649         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 650         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 651         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 652         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 653         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 654         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 655         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 656         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 660         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 661         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 662         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 663         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 664         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 665         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 666         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 667         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 668         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 669         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 670         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 671         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 672         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 673         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 674         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 675         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 676         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 677         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 678         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 679         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 680         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 681         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 682         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 683         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 684         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 685         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 686         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 687         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 688         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 690         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 692         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 693         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 694         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 695         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 696         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 697         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 698         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 699         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 700         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 701         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 702         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 703         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 704         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 705         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 706         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 707 };
 708
 709 static const struct nla_policy
 710 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 711         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 712 };
 713
 714 static const struct nla_policy
 715 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 716         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 717         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 718         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 719                                                        .len = 128 },
 720 };
 721
 722 static void fl_set_key_val(struct nlattr **tb,
 723                            void *val, int val_type,
 724                            void *mask, int mask_type, int len)
 725 {
 726         if (!tb[val_type])
 727                 return;
 728         memcpy(val, nla_data(tb[val_type]), len);
 729         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 730                 memset(mask, 0xff, len);
 731         else
 732                 memcpy(mask, nla_data(tb[mask_type]), len);
 733 }
 734
 735 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 736                                  struct fl_flow_key *mask)
 737 {
 738         fl_set_key_val(tb, &key->tp_min.dst,
 739                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_min.dst,
 740                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.dst));
 741         fl_set_key_val(tb, &key->tp_max.dst,
 742                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_max.dst,
 743                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.dst));
 744         fl_set_key_val(tb, &key->tp_min.src,
 745                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_min.src,
 746                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.src));
 747         fl_set_key_val(tb, &key->tp_max.src,
 748                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_max.src,
 749                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.src));
 750
 751         if ((mask->tp_min.dst && mask->tp_max.dst &&
 752              htons(key->tp_max.dst) <= htons(key->tp_min.dst)) ||
 753              (mask->tp_min.src && mask->tp_max.src &&
 754               htons(key->tp_max.src) <= htons(key->tp_min.src)))
 755                 return -EINVAL;
 756
 757         return 0;
 758 }
 759
 760 static int fl_set_key_mpls(struct nlattr **tb,
 761                            struct flow_dissector_key_mpls *key_val,
 762                            struct flow_dissector_key_mpls *key_mask)
 763 {
 764         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 765                 key_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 766                 key_mask->mpls_ttl = MPLS_TTL_MASK;
 767         }
 768         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 769                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 770
 771                 if (bos & ~MPLS_BOS_MASK)
 772                         return -EINVAL;
 773                 key_val->mpls_bos = bos;
 774                 key_mask->mpls_bos = MPLS_BOS_MASK;
 775         }
 776         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 777                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 778
 779                 if (tc & ~MPLS_TC_MASK)
 780                         return -EINVAL;
 781                 key_val->mpls_tc = tc;
 782                 key_mask->mpls_tc = MPLS_TC_MASK;
 783         }
 784         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 785                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 786
 787                 if (label & ~MPLS_LABEL_MASK)
 788                         return -EINVAL;
 789                 key_val->mpls_label = label;
 790                 key_mask->mpls_label = MPLS_LABEL_MASK;
 791         }
 792         return 0;
 793 }
 794
 795 static void fl_set_key_vlan(struct nlattr **tb,
 796                             __be16 ethertype,
 797                             int vlan_id_key, int vlan_prio_key,
 798                             struct flow_dissector_key_vlan *key_val,
 799                             struct flow_dissector_key_vlan *key_mask)
 800 {
 801 #define VLAN_PRIORITY_MASK      0x7
 802
 803         if (tb[vlan_id_key]) {
 804                 key_val->vlan_id =
 805                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
 806                 key_mask->vlan_id = VLAN_VID_MASK;
 807         }
 808         if (tb[vlan_prio_key]) {
 809                 key_val->vlan_priority =
 810                         nla_get_u8(tb[vlan_prio_key]) &
 811                         VLAN_PRIORITY_MASK;
 812                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 813         }
 814         key_val->vlan_tpid = ethertype;
 815         key_mask->vlan_tpid = cpu_to_be16(~0);
 816 }
 817
 818 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
 819                             u32 *dissector_key, u32 *dissector_mask,
 820                             u32 flower_flag_bit, u32 dissector_flag_bit)
 821 {
 822         if (flower_mask & flower_flag_bit) {
 823                 *dissector_mask |= dissector_flag_bit;
 824                 if (flower_key & flower_flag_bit)
 825                         *dissector_key |= dissector_flag_bit;
 826         }
 827 }
 828
 829 static int fl_set_key_flags(struct nlattr **tb,
 830                             u32 *flags_key, u32 *flags_mask)
 831 {
 832         u32 key, mask;
 833
 834         /* mask is mandatory for flags */
 835         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK])
 836                 return -EINVAL;
 837
 838         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
 839         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
 840
 841         *flags_key  = 0;
 842         *flags_mask = 0;
 843
 844         fl_set_key_flag(key, mask, flags_key, flags_mask,
 845                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
 846         fl_set_key_flag(key, mask, flags_key, flags_mask,
 847                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
 848                         FLOW_DIS_FIRST_FRAG);
 849
 850         return 0;
 851 }
 852
 853 static void fl_set_key_ip(struct nlattr **tb, bool encap,
 854                           struct flow_dissector_key_ip *key,
 855                           struct flow_dissector_key_ip *mask)
 856 {
 857         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
 858         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
 859         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
 860         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
 861
 862         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
 863         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
 864 }
 865
 866 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
 867                              int depth, int option_len,
 868                              struct netlink_ext_ack *extack)
 869 {
 870         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
 871         struct nlattr *class = NULL, *type = NULL, *data = NULL;
 872         struct geneve_opt *opt;
 873         int err, data_len = 0;
 874
 875         if (option_len > sizeof(struct geneve_opt))
 876                 data_len = option_len - sizeof(struct geneve_opt);
 877
 878         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
 879         memset(opt, 0xff, option_len);
 880         opt->length = data_len / 4;
 881         opt->r1 = 0;
 882         opt->r2 = 0;
 883         opt->r3 = 0;
 884
 885         /* If no mask has been prodived we assume an exact match. */
 886         if (!depth)
 887                 return sizeof(struct geneve_opt) + data_len;
 888
 889         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
 890                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
 891                 return -EINVAL;
 892         }
 893
 894         err = nla_parse_nested_deprecated(tb,
 895                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
 896                                           nla, geneve_opt_policy, extack);
 897         if (err < 0)
 898                 return err;
 899
 900         /* We are not allowed to omit any of CLASS, TYPE or DATA
 901          * fields from the key.
 902          */
 903         if (!option_len &&
 904             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
 905              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
 906              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
 907                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
 908                 return -EINVAL;
 909         }
 910
 911         /* Omitting any of CLASS, TYPE or DATA fields is allowed
 912          * for the mask.
 913          */
 914         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
 915                 int new_len = key->enc_opts.len;
 916
 917                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
 918                 data_len = nla_len(data);
 919                 if (data_len < 4) {
 920                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
 921                         return -ERANGE;
 922                 }
 923                 if (data_len % 4) {
 924                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
 925                         return -ERANGE;
 926                 }
 927
 928                 new_len += sizeof(struct geneve_opt) + data_len;
 929                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
 930                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
 931                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
 932                         return -ERANGE;
 933                 }
 934                 opt->length = data_len / 4;
 935                 memcpy(opt->opt_data, nla_data(data), data_len);
 936         }
 937
 938         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
 939                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
 940                 opt->opt_class = nla_get_be16(class);
 941         }
 942
 943         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
 944                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
 945                 opt->type = nla_get_u8(type);
 946         }
 947
 948         return sizeof(struct geneve_opt) + data_len;
 949 }
 950
 951 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
 952                           struct fl_flow_key *mask,
 953                           struct netlink_ext_ack *extack)
 954 {
 955         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
 956         int err, option_len, key_depth, msk_depth = 0;
 957
 958         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
 959                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
 960                                              enc_opts_policy, extack);
 961         if (err)
 962                 return err;
 963
 964         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
 965
 966         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
 967                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
 968                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
 969                                                      enc_opts_policy, extack);
 970                 if (err)
 971                         return err;
 972
 973                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 974                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 975         }
 976
 977         nla_for_each_attr(nla_opt_key, nla_enc_key,
 978                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
 979                 switch (nla_type(nla_opt_key)) {
 980                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
 981                         option_len = 0;
 982                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 983                         option_len = fl_set_geneve_opt(nla_opt_key, key,
 984                                                        key_depth, option_len,
 985                                                        extack);
 986                         if (option_len < 0)
 987                                 return option_len;
 988
 989                         key->enc_opts.len += option_len;
 990                         /* At the same time we need to parse through the mask
 991                          * in order to verify exact and mask attribute lengths.
 992                          */
 993                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 994                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
 995                                                        msk_depth, option_len,
 996                                                        extack);
 997                         if (option_len < 0)
 998                                 return option_len;
 999
1000                         mask->enc_opts.len += option_len;
1001                         if (key->enc_opts.len != mask->enc_opts.len) {
1002                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1003                                 return -EINVAL;
1004                         }
1005
1006                         if (msk_depth)
1007                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1008                         break;
1009                 default:
1010                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1011                         return -EINVAL;
1012                 }
1013         }
1014
1015         return 0;
1016 }
1017
1018 static int fl_set_key(struct net *net, struct nlattr **tb,
1019                       struct fl_flow_key *key, struct fl_flow_key *mask,
1020                       struct netlink_ext_ack *extack)
1021 {
1022         __be16 ethertype;
1023         int ret = 0;
1024 #ifdef CONFIG_NET_CLS_IND
1025         if (tb[TCA_FLOWER_INDEV]) {
1026                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1027                 if (err < 0)
1028                         return err;
1029                 key->indev_ifindex = err;
1030                 mask->indev_ifindex = 0xffffffff;
1031         }
1032 #endif
1033
1034         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1035                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1036                        sizeof(key->eth.dst));
1037         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1038                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1039                        sizeof(key->eth.src));
1040
1041         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1042                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1043
1044                 if (eth_type_vlan(ethertype)) {
1045                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1046                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
1047                                         &mask->vlan);
1048
1049                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1050                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1051                                 if (eth_type_vlan(ethertype)) {
1052                                         fl_set_key_vlan(tb, ethertype,
1053                                                         TCA_FLOWER_KEY_CVLAN_ID,
1054                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1055                                                         &key->cvlan, &mask->cvlan);
1056                                         fl_set_key_val(tb, &key->basic.n_proto,
1057                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1058                                                        &mask->basic.n_proto,
1059                                                        TCA_FLOWER_UNSPEC,
1060                                                        sizeof(key->basic.n_proto));
1061                                 } else {
1062                                         key->basic.n_proto = ethertype;
1063                                         mask->basic.n_proto = cpu_to_be16(~0);
1064                                 }
1065                         }
1066                 } else {
1067                         key->basic.n_proto = ethertype;
1068                         mask->basic.n_proto = cpu_to_be16(~0);
1069                 }
1070         }
1071
1072         if (key->basic.n_proto == htons(ETH_P_IP) ||
1073             key->basic.n_proto == htons(ETH_P_IPV6)) {
1074                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1075                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1076                                sizeof(key->basic.ip_proto));
1077                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1078         }
1079
1080         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1081                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1082                 mask->control.addr_type = ~0;
1083                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1084                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1085                                sizeof(key->ipv4.src));
1086                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1087                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1088                                sizeof(key->ipv4.dst));
1089         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1090                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1091                 mask->control.addr_type = ~0;
1092                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1093                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1094                                sizeof(key->ipv6.src));
1095                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1096                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1097                                sizeof(key->ipv6.dst));
1098         }
1099
1100         if (key->basic.ip_proto == IPPROTO_TCP) {
1101                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1102                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1103                                sizeof(key->tp.src));
1104                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1105                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1106                                sizeof(key->tp.dst));
1107                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1108                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1109                                sizeof(key->tcp.flags));
1110         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1111                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1112                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1113                                sizeof(key->tp.src));
1114                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1115                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1116                                sizeof(key->tp.dst));
1117         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1118                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1119                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1120                                sizeof(key->tp.src));
1121                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1122                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1123                                sizeof(key->tp.dst));
1124         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1125                    key->basic.ip_proto == IPPROTO_ICMP) {
1126                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1127                                &mask->icmp.type,
1128                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1129                                sizeof(key->icmp.type));
1130                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1131                                &mask->icmp.code,
1132                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1133                                sizeof(key->icmp.code));
1134         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1135                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1136                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1137                                &mask->icmp.type,
1138                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1139                                sizeof(key->icmp.type));
1140                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1141                                &mask->icmp.code,
1142                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1143                                sizeof(key->icmp.code));
1144         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1145                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1146                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls);
1147                 if (ret)
1148                         return ret;
1149         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1150                    key->basic.n_proto == htons(ETH_P_RARP)) {
1151                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1152                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1153                                sizeof(key->arp.sip));
1154                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1155                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1156                                sizeof(key->arp.tip));
1157                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1158                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1159                                sizeof(key->arp.op));
1160                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1161                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1162                                sizeof(key->arp.sha));
1163                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1164                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1165                                sizeof(key->arp.tha));
1166         }
1167
1168         if (key->basic.ip_proto == IPPROTO_TCP ||
1169             key->basic.ip_proto == IPPROTO_UDP ||
1170             key->basic.ip_proto == IPPROTO_SCTP) {
1171                 ret = fl_set_key_port_range(tb, key, mask);
1172                 if (ret)
1173                         return ret;
1174         }
1175
1176         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1177             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1178                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1179                 mask->enc_control.addr_type = ~0;
1180                 fl_set_key_val(tb, &key->enc_ipv4.src,
1181                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1182                                &mask->enc_ipv4.src,
1183                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1184                                sizeof(key->enc_ipv4.src));
1185                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1186                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1187                                &mask->enc_ipv4.dst,
1188                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1189                                sizeof(key->enc_ipv4.dst));
1190         }
1191
1192         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1193             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1194                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1195                 mask->enc_control.addr_type = ~0;
1196                 fl_set_key_val(tb, &key->enc_ipv6.src,
1197                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1198                                &mask->enc_ipv6.src,
1199                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1200                                sizeof(key->enc_ipv6.src));
1201                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1202                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1203                                &mask->enc_ipv6.dst,
1204                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1205                                sizeof(key->enc_ipv6.dst));
1206         }
1207
1208         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1209                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1210                        sizeof(key->enc_key_id.keyid));
1211
1212         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1213                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1214                        sizeof(key->enc_tp.src));
1215
1216         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1217                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1218                        sizeof(key->enc_tp.dst));
1219
1220         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1221
1222         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1223                 ret = fl_set_enc_opt(tb, key, mask, extack);
1224                 if (ret)
1225                         return ret;
1226         }
1227
1228         if (tb[TCA_FLOWER_KEY_FLAGS])
1229                 ret = fl_set_key_flags(tb, &key->control.flags, &mask->control.flags);
1230
1231         return ret;
1232 }
1233
1234 static void fl_mask_copy(struct fl_flow_mask *dst,
1235                          struct fl_flow_mask *src)
1236 {
1237         const void *psrc = fl_key_get_start(&src->key, src);
1238         void *pdst = fl_key_get_start(&dst->key, src);
1239
1240         memcpy(pdst, psrc, fl_mask_range(src));
1241         dst->range = src->range;
1242 }
1243
1244 static const struct rhashtable_params fl_ht_params = {
1245         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1246         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1247         .automatic_shrinking = true,
1248 };
1249
1250 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1251 {
1252         mask->filter_ht_params = fl_ht_params;
1253         mask->filter_ht_params.key_len = fl_mask_range(mask);
1254         mask->filter_ht_params.key_offset += mask->range.start;
1255
1256         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1257 }
1258
1259 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1260 #define FL_KEY_MEMBER_SIZE(member) FIELD_SIZEOF(struct fl_flow_key, member)
1261
1262 #define FL_KEY_IS_MASKED(mask, member)                                          \
1263         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1264                    0, FL_KEY_MEMBER_SIZE(member))                               \
1265
1266 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1267         do {                                                                    \
1268                 keys[cnt].key_id = id;                                          \
1269                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1270                 cnt++;                                                          \
1271         } while(0);
1272
1273 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1274         do {                                                                    \
1275                 if (FL_KEY_IS_MASKED(mask, member))                             \
1276                         FL_KEY_SET(keys, cnt, id, member);                      \
1277         } while(0);
1278
1279 static void fl_init_dissector(struct flow_dissector *dissector,
1280                               struct fl_flow_key *mask)
1281 {
1282         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1283         size_t cnt = 0;
1284
1285         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1286         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1287         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1288                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1289         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1290                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1291         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1292                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1293         if (FL_KEY_IS_MASKED(mask, tp) ||
1294             FL_KEY_IS_MASKED(mask, tp_min) || FL_KEY_IS_MASKED(mask, tp_max))
1295                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_PORTS, tp);
1296         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1297                              FLOW_DISSECTOR_KEY_IP, ip);
1298         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1299                              FLOW_DISSECTOR_KEY_TCP, tcp);
1300         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1301                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1302         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1303                              FLOW_DISSECTOR_KEY_ARP, arp);
1304         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1305                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1306         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1307                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1308         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1309                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1310         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1311                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1312         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1313                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1314         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1315                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1316         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1317             FL_KEY_IS_MASKED(mask, enc_ipv6))
1318                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1319                            enc_control);
1320         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1321                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1322         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1323                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1324         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1325                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1326
1327         skb_flow_dissector_init(dissector, keys, cnt);
1328 }
1329
1330 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1331                                                struct fl_flow_mask *mask)
1332 {
1333         struct fl_flow_mask *newmask;
1334         int err;
1335
1336         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1337         if (!newmask)
1338                 return ERR_PTR(-ENOMEM);
1339
1340         fl_mask_copy(newmask, mask);
1341
1342         if ((newmask->key.tp_min.dst && newmask->key.tp_max.dst) ||
1343             (newmask->key.tp_min.src && newmask->key.tp_max.src))
1344                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1345
1346         err = fl_init_mask_hashtable(newmask);
1347         if (err)
1348                 goto errout_free;
1349
1350         fl_init_dissector(&newmask->dissector, &newmask->key);
1351
1352         INIT_LIST_HEAD_RCU(&newmask->filters);
1353
1354         refcount_set(&newmask->refcnt, 1);
1355         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1356                                       &newmask->ht_node, mask_ht_params);
1357         if (err)
1358                 goto errout_destroy;
1359
1360         spin_lock(&head->masks_lock);
1361         list_add_tail_rcu(&newmask->list, &head->masks);
1362         spin_unlock(&head->masks_lock);
1363
1364         return newmask;
1365
1366 errout_destroy:
1367         rhashtable_destroy(&newmask->ht);
1368 errout_free:
1369         kfree(newmask);
1370
1371         return ERR_PTR(err);
1372 }
1373
1374 static int fl_check_assign_mask(struct cls_fl_head *head,
1375                                 struct cls_fl_filter *fnew,
1376                                 struct cls_fl_filter *fold,
1377                                 struct fl_flow_mask *mask)
1378 {
1379         struct fl_flow_mask *newmask;
1380         int ret = 0;
1381
1382         rcu_read_lock();
1383
1384         /* Insert mask as temporary node to prevent concurrent creation of mask
1385          * with same key. Any concurrent lookups with same key will return
1386          * -EAGAIN because mask's refcnt is zero.
1387          */
1388         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1389                                                        &mask->ht_node,
1390                                                        mask_ht_params);
1391         if (!fnew->mask) {
1392                 rcu_read_unlock();
1393
1394                 if (fold) {
1395                         ret = -EINVAL;
1396                         goto errout_cleanup;
1397                 }
1398
1399                 newmask = fl_create_new_mask(head, mask);
1400                 if (IS_ERR(newmask)) {
1401                         ret = PTR_ERR(newmask);
1402                         goto errout_cleanup;
1403                 }
1404
1405                 fnew->mask = newmask;
1406                 return 0;
1407         } else if (IS_ERR(fnew->mask)) {
1408                 ret = PTR_ERR(fnew->mask);
1409         } else if (fold && fold->mask != fnew->mask) {
1410                 ret = -EINVAL;
1411         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1412                 /* Mask was deleted concurrently, try again */
1413                 ret = -EAGAIN;
1414         }
1415         rcu_read_unlock();
1416         return ret;
1417
1418 errout_cleanup:
1419         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1420                                mask_ht_params);
1421         return ret;
1422 }
1423
1424 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1425                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1426                         unsigned long base, struct nlattr **tb,
1427                         struct nlattr *est, bool ovr,
1428                         struct fl_flow_tmplt *tmplt, bool rtnl_held,
1429                         struct netlink_ext_ack *extack)
1430 {
1431         int err;
1432
1433         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, rtnl_held,
1434                                 extack);
1435         if (err < 0)
1436                 return err;
1437
1438         if (tb[TCA_FLOWER_CLASSID]) {
1439                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1440                 if (!rtnl_held)
1441                         rtnl_lock();
1442                 tcf_bind_filter(tp, &f->res, base);
1443                 if (!rtnl_held)
1444                         rtnl_unlock();
1445         }
1446
1447         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1448         if (err)
1449                 return err;
1450
1451         fl_mask_update_range(mask);
1452         fl_set_masked_key(&f->mkey, &f->key, mask);
1453
1454         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1455                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1456                 return -EINVAL;
1457         }
1458
1459         return 0;
1460 }
1461
1462 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
1463                                struct cls_fl_filter *fold,
1464                                bool *in_ht)
1465 {
1466         struct fl_flow_mask *mask = fnew->mask;
1467         int err;
1468
1469         err = rhashtable_lookup_insert_fast(&mask->ht,
1470                                             &fnew->ht_node,
1471                                             mask->filter_ht_params);
1472         if (err) {
1473                 *in_ht = false;
1474                 /* It is okay if filter with same key exists when
1475                  * overwriting.
1476                  */
1477                 return fold && err == -EEXIST ? 0 : err;
1478         }
1479
1480         *in_ht = true;
1481         return 0;
1482 }
1483
1484 static int fl_change(struct net *net, struct sk_buff *in_skb,
1485                      struct tcf_proto *tp, unsigned long base,
1486                      u32 handle, struct nlattr **tca,
1487                      void **arg, bool ovr, bool rtnl_held,
1488                      struct netlink_ext_ack *extack)
1489 {
1490         struct cls_fl_head *head = fl_head_dereference(tp);
1491         struct cls_fl_filter *fold = *arg;
1492         struct cls_fl_filter *fnew;
1493         struct fl_flow_mask *mask;
1494         struct nlattr **tb;
1495         bool in_ht;
1496         int err;
1497
1498         if (!tca[TCA_OPTIONS]) {
1499                 err = -EINVAL;
1500                 goto errout_fold;
1501         }
1502
1503         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
1504         if (!mask) {
1505                 err = -ENOBUFS;
1506                 goto errout_fold;
1507         }
1508
1509         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1510         if (!tb) {
1511                 err = -ENOBUFS;
1512                 goto errout_mask_alloc;
1513         }
1514
1515         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1516                                           tca[TCA_OPTIONS], fl_policy, NULL);
1517         if (err < 0)
1518                 goto errout_tb;
1519
1520         if (fold && handle && fold->handle != handle) {
1521                 err = -EINVAL;
1522                 goto errout_tb;
1523         }
1524
1525         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
1526         if (!fnew) {
1527                 err = -ENOBUFS;
1528                 goto errout_tb;
1529         }
1530         INIT_LIST_HEAD(&fnew->hw_list);
1531         refcount_set(&fnew->refcnt, 1);
1532
1533         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
1534         if (err < 0)
1535                 goto errout;
1536
1537         if (tb[TCA_FLOWER_FLAGS]) {
1538                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
1539
1540                 if (!tc_flags_valid(fnew->flags)) {
1541                         err = -EINVAL;
1542                         goto errout;
1543                 }
1544         }
1545
1546         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE], ovr,
1547                            tp->chain->tmplt_priv, rtnl_held, extack);
1548         if (err)
1549                 goto errout;
1550
1551         err = fl_check_assign_mask(head, fnew, fold, mask);
1552         if (err)
1553                 goto errout;
1554
1555         err = fl_ht_insert_unique(fnew, fold, &in_ht);
1556         if (err)
1557                 goto errout_mask;
1558
1559         if (!tc_skip_hw(fnew->flags)) {
1560                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
1561                 if (err)
1562                         goto errout_ht;
1563         }
1564
1565         if (!tc_in_hw(fnew->flags))
1566                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1567
1568         spin_lock(&tp->lock);
1569
1570         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
1571          * proto again or create new one, if necessary.
1572          */
1573         if (tp->deleting) {
1574                 err = -EAGAIN;
1575                 goto errout_hw;
1576         }
1577
1578         if (fold) {
1579                 /* Fold filter was deleted concurrently. Retry lookup. */
1580                 if (fold->deleted) {
1581                         err = -EAGAIN;
1582                         goto errout_hw;
1583                 }
1584
1585                 fnew->handle = handle;
1586
1587                 if (!in_ht) {
1588                         struct rhashtable_params params =
1589                                 fnew->mask->filter_ht_params;
1590
1591                         err = rhashtable_insert_fast(&fnew->mask->ht,
1592                                                      &fnew->ht_node,
1593                                                      params);
1594                         if (err)
1595                                 goto errout_hw;
1596                         in_ht = true;
1597                 }
1598
1599                 refcount_inc(&fnew->refcnt);
1600                 rhashtable_remove_fast(&fold->mask->ht,
1601                                        &fold->ht_node,
1602                                        fold->mask->filter_ht_params);
1603                 idr_replace(&head->handle_idr, fnew, fnew->handle);
1604                 list_replace_rcu(&fold->list, &fnew->list);
1605                 fold->deleted = true;
1606
1607                 spin_unlock(&tp->lock);
1608
1609                 fl_mask_put(head, fold->mask);
1610                 if (!tc_skip_hw(fold->flags))
1611                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
1612                 tcf_unbind_filter(tp, &fold->res);
1613                 /* Caller holds reference to fold, so refcnt is always > 0
1614                  * after this.
1615                  */
1616                 refcount_dec(&fold->refcnt);
1617                 __fl_put(fold);
1618         } else {
1619                 if (handle) {
1620                         /* user specifies a handle and it doesn't exist */
1621                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1622                                             handle, GFP_ATOMIC);
1623
1624                         /* Filter with specified handle was concurrently
1625                          * inserted after initial check in cls_api. This is not
1626                          * necessarily an error if NLM_F_EXCL is not set in
1627                          * message flags. Returning EAGAIN will cause cls_api to
1628                          * try to update concurrently inserted rule.
1629                          */
1630                         if (err == -ENOSPC)
1631                                 err = -EAGAIN;
1632                 } else {
1633                         handle = 1;
1634                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1635                                             INT_MAX, GFP_ATOMIC);
1636                 }
1637                 if (err)
1638                         goto errout_hw;
1639
1640                 refcount_inc(&fnew->refcnt);
1641                 fnew->handle = handle;
1642                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
1643                 spin_unlock(&tp->lock);
1644         }
1645
1646         *arg = fnew;
1647
1648         kfree(tb);
1649         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
1650         return 0;
1651
1652 errout_ht:
1653         spin_lock(&tp->lock);
1654 errout_hw:
1655         fnew->deleted = true;
1656         spin_unlock(&tp->lock);
1657         if (!tc_skip_hw(fnew->flags))
1658                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
1659         if (in_ht)
1660                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
1661                                        fnew->mask->filter_ht_params);
1662 errout_mask:
1663         fl_mask_put(head, fnew->mask);
1664 errout:
1665         __fl_put(fnew);
1666 errout_tb:
1667         kfree(tb);
1668 errout_mask_alloc:
1669         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
1670 errout_fold:
1671         if (fold)
1672                 __fl_put(fold);
1673         return err;
1674 }
1675
1676 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
1677                      bool rtnl_held, struct netlink_ext_ack *extack)
1678 {
1679         struct cls_fl_head *head = fl_head_dereference(tp);
1680         struct cls_fl_filter *f = arg;
1681         bool last_on_mask;
1682         int err = 0;
1683
1684         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
1685         *last = list_empty(&head->masks);
1686         __fl_put(f);
1687
1688         return err;
1689 }
1690
1691 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
1692                     bool rtnl_held)
1693 {
1694         struct cls_fl_filter *f;
1695
1696         arg->count = arg->skip;
1697
1698         while ((f = fl_get_next_filter(tp, &arg->cookie)) != NULL) {
1699                 if (arg->fn(tp, f, arg) < 0) {
1700                         __fl_put(f);
1701                         arg->stop = 1;
1702                         break;
1703                 }
1704                 __fl_put(f);
1705                 arg->cookie++;
1706                 arg->count++;
1707         }
1708 }
1709
1710 static struct cls_fl_filter *
1711 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
1712 {
1713         struct cls_fl_head *head = fl_head_dereference(tp);
1714
1715         spin_lock(&tp->lock);
1716         if (list_empty(&head->hw_filters)) {
1717                 spin_unlock(&tp->lock);
1718                 return NULL;
1719         }
1720
1721         if (!f)
1722                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
1723                                hw_list);
1724         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
1725                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
1726                         spin_unlock(&tp->lock);
1727                         return f;
1728                 }
1729         }
1730
1731         spin_unlock(&tp->lock);
1732         return NULL;
1733 }
1734
1735 static int fl_reoffload(struct tcf_proto *tp, bool add, tc_setup_cb_t *cb,
1736                         void *cb_priv, struct netlink_ext_ack *extack)
1737 {
1738         struct tc_cls_flower_offload cls_flower = {};
1739         struct tcf_block *block = tp->chain->block;
1740         struct cls_fl_filter *f = NULL;
1741         int err;
1742
1743         /* hw_filters list can only be changed by hw offload functions after
1744          * obtaining rtnl lock. Make sure it is not changed while reoffload is
1745          * iterating it.
1746          */
1747         ASSERT_RTNL();
1748
1749         while ((f = fl_get_next_hw_filter(tp, f, add))) {
1750                 cls_flower.rule =
1751                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
1752                 if (!cls_flower.rule) {
1753                         __fl_put(f);
1754                         return -ENOMEM;
1755                 }
1756
1757                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
1758                                            extack);
1759                 cls_flower.command = add ?
1760                         TC_CLSFLOWER_REPLACE : TC_CLSFLOWER_DESTROY;
1761                 cls_flower.cookie = (unsigned long)f;
1762                 cls_flower.rule->match.dissector = &f->mask->dissector;
1763                 cls_flower.rule->match.mask = &f->mask->key;
1764                 cls_flower.rule->match.key = &f->mkey;
1765
1766                 err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
1767                 if (err) {
1768                         kfree(cls_flower.rule);
1769                         if (tc_skip_sw(f->flags)) {
1770                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
1771                                 __fl_put(f);
1772                                 return err;
1773                         }
1774                         goto next_flow;
1775                 }
1776
1777                 cls_flower.classid = f->res.classid;
1778
1779                 err = cb(TC_SETUP_CLSFLOWER, &cls_flower, cb_priv);
1780                 kfree(cls_flower.rule);
1781
1782                 if (err) {
1783                         if (add && tc_skip_sw(f->flags)) {
1784                                 __fl_put(f);
1785                                 return err;
1786                         }
1787                         goto next_flow;
1788                 }
1789
1790                 spin_lock(&tp->lock);
1791                 tc_cls_offload_cnt_update(block, &f->in_hw_count, &f->flags,
1792                                           add);
1793                 spin_unlock(&tp->lock);
1794 next_flow:
1795                 __fl_put(f);
1796         }
1797
1798         return 0;
1799 }
1800
1801 static int fl_hw_create_tmplt(struct tcf_chain *chain,
1802                               struct fl_flow_tmplt *tmplt)
1803 {
1804         struct tc_cls_flower_offload cls_flower = {};
1805         struct tcf_block *block = chain->block;
1806
1807         cls_flower.rule = flow_rule_alloc(0);
1808         if (!cls_flower.rule)
1809                 return -ENOMEM;
1810
1811         cls_flower.common.chain_index = chain->index;
1812         cls_flower.command = TC_CLSFLOWER_TMPLT_CREATE;
1813         cls_flower.cookie = (unsigned long) tmplt;
1814         cls_flower.rule->match.dissector = &tmplt->dissector;
1815         cls_flower.rule->match.mask = &tmplt->mask;
1816         cls_flower.rule->match.key = &tmplt->dummy_key;
1817
1818         /* We don't care if driver (any of them) fails to handle this
1819          * call. It serves just as a hint for it.
1820          */
1821         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
1822         kfree(cls_flower.rule);
1823
1824         return 0;
1825 }
1826
1827 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
1828                                 struct fl_flow_tmplt *tmplt)
1829 {
1830         struct tc_cls_flower_offload cls_flower = {};
1831         struct tcf_block *block = chain->block;
1832
1833         cls_flower.common.chain_index = chain->index;
1834         cls_flower.command = TC_CLSFLOWER_TMPLT_DESTROY;
1835         cls_flower.cookie = (unsigned long) tmplt;
1836
1837         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
1838 }
1839
1840 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
1841                              struct nlattr **tca,
1842                              struct netlink_ext_ack *extack)
1843 {
1844         struct fl_flow_tmplt *tmplt;
1845         struct nlattr **tb;
1846         int err;
1847
1848         if (!tca[TCA_OPTIONS])
1849                 return ERR_PTR(-EINVAL);
1850
1851         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1852         if (!tb)
1853                 return ERR_PTR(-ENOBUFS);
1854         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1855                                           tca[TCA_OPTIONS], fl_policy, NULL);
1856         if (err)
1857                 goto errout_tb;
1858
1859         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
1860         if (!tmplt) {
1861                 err = -ENOMEM;
1862                 goto errout_tb;
1863         }
1864         tmplt->chain = chain;
1865         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
1866         if (err)
1867                 goto errout_tmplt;
1868
1869         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
1870
1871         err = fl_hw_create_tmplt(chain, tmplt);
1872         if (err)
1873                 goto errout_tmplt;
1874
1875         kfree(tb);
1876         return tmplt;
1877
1878 errout_tmplt:
1879         kfree(tmplt);
1880 errout_tb:
1881         kfree(tb);
1882         return ERR_PTR(err);
1883 }
1884
1885 static void fl_tmplt_destroy(void *tmplt_priv)
1886 {
1887         struct fl_flow_tmplt *tmplt = tmplt_priv;
1888
1889         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
1890         kfree(tmplt);
1891 }
1892
1893 static int fl_dump_key_val(struct sk_buff *skb,
1894                            void *val, int val_type,
1895                            void *mask, int mask_type, int len)
1896 {
1897         int err;
1898
1899         if (!memchr_inv(mask, 0, len))
1900                 return 0;
1901         err = nla_put(skb, val_type, len, val);
1902         if (err)
1903                 return err;
1904         if (mask_type != TCA_FLOWER_UNSPEC) {
1905                 err = nla_put(skb, mask_type, len, mask);
1906                 if (err)
1907                         return err;
1908         }
1909         return 0;
1910 }
1911
1912 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
1913                                   struct fl_flow_key *mask)
1914 {
1915         if (fl_dump_key_val(skb, &key->tp_min.dst, TCA_FLOWER_KEY_PORT_DST_MIN,
1916                             &mask->tp_min.dst, TCA_FLOWER_UNSPEC,
1917                             sizeof(key->tp_min.dst)) ||
1918             fl_dump_key_val(skb, &key->tp_max.dst, TCA_FLOWER_KEY_PORT_DST_MAX,
1919                             &mask->tp_max.dst, TCA_FLOWER_UNSPEC,
1920                             sizeof(key->tp_max.dst)) ||
1921             fl_dump_key_val(skb, &key->tp_min.src, TCA_FLOWER_KEY_PORT_SRC_MIN,
1922                             &mask->tp_min.src, TCA_FLOWER_UNSPEC,
1923                             sizeof(key->tp_min.src)) ||
1924             fl_dump_key_val(skb, &key->tp_max.src, TCA_FLOWER_KEY_PORT_SRC_MAX,
1925                             &mask->tp_max.src, TCA_FLOWER_UNSPEC,
1926                             sizeof(key->tp_max.src)))
1927                 return -1;
1928
1929         return 0;
1930 }
1931
1932 static int fl_dump_key_mpls(struct sk_buff *skb,
1933                             struct flow_dissector_key_mpls *mpls_key,
1934                             struct flow_dissector_key_mpls *mpls_mask)
1935 {
1936         int err;
1937
1938         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
1939                 return 0;
1940         if (mpls_mask->mpls_ttl) {
1941                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
1942                                  mpls_key->mpls_ttl);
1943                 if (err)
1944                         return err;
1945         }
1946         if (mpls_mask->mpls_tc) {
1947                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
1948                                  mpls_key->mpls_tc);
1949                 if (err)
1950                         return err;
1951         }
1952         if (mpls_mask->mpls_label) {
1953                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
1954                                   mpls_key->mpls_label);
1955                 if (err)
1956                         return err;
1957         }
1958         if (mpls_mask->mpls_bos) {
1959                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
1960                                  mpls_key->mpls_bos);
1961                 if (err)
1962                         return err;
1963         }
1964         return 0;
1965 }
1966
1967 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
1968                           struct flow_dissector_key_ip *key,
1969                           struct flow_dissector_key_ip *mask)
1970 {
1971         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1972         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1973         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1974         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1975
1976         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
1977             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
1978                 return -1;
1979
1980         return 0;
1981 }
1982
1983 static int fl_dump_key_vlan(struct sk_buff *skb,
1984                             int vlan_id_key, int vlan_prio_key,
1985                             struct flow_dissector_key_vlan *vlan_key,
1986                             struct flow_dissector_key_vlan *vlan_mask)
1987 {
1988         int err;
1989
1990         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
1991                 return 0;
1992         if (vlan_mask->vlan_id) {
1993                 err = nla_put_u16(skb, vlan_id_key,
1994                                   vlan_key->vlan_id);
1995                 if (err)
1996                         return err;
1997         }
1998         if (vlan_mask->vlan_priority) {
1999                 err = nla_put_u8(skb, vlan_prio_key,
2000                                  vlan_key->vlan_priority);
2001                 if (err)
2002                         return err;
2003         }
2004         return 0;
2005 }
2006
2007 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2008                             u32 *flower_key, u32 *flower_mask,
2009                             u32 flower_flag_bit, u32 dissector_flag_bit)
2010 {
2011         if (dissector_mask & dissector_flag_bit) {
2012                 *flower_mask |= flower_flag_bit;
2013                 if (dissector_key & dissector_flag_bit)
2014                         *flower_key |= flower_flag_bit;
2015         }
2016 }
2017
2018 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2019 {
2020         u32 key, mask;
2021         __be32 _key, _mask;
2022         int err;
2023
2024         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2025                 return 0;
2026
2027         key = 0;
2028         mask = 0;
2029
2030         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2031                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2032         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2033                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2034                         FLOW_DIS_FIRST_FRAG);
2035
2036         _key = cpu_to_be32(key);
2037         _mask = cpu_to_be32(mask);
2038
2039         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2040         if (err)
2041                 return err;
2042
2043         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2044 }
2045
2046 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2047                                   struct flow_dissector_key_enc_opts *enc_opts)
2048 {
2049         struct geneve_opt *opt;
2050         struct nlattr *nest;
2051         int opt_off = 0;
2052
2053         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2054         if (!nest)
2055                 goto nla_put_failure;
2056
2057         while (enc_opts->len > opt_off) {
2058                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2059
2060                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2061                                  opt->opt_class))
2062                         goto nla_put_failure;
2063                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2064                                opt->type))
2065                         goto nla_put_failure;
2066                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2067                             opt->length * 4, opt->opt_data))
2068                         goto nla_put_failure;
2069
2070                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2071         }
2072         nla_nest_end(skb, nest);
2073         return 0;
2074
2075 nla_put_failure:
2076         nla_nest_cancel(skb, nest);
2077         return -EMSGSIZE;
2078 }
2079
2080 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2081                                struct flow_dissector_key_enc_opts *enc_opts)
2082 {
2083         struct nlattr *nest;
2084         int err;
2085
2086         if (!enc_opts->len)
2087                 return 0;
2088
2089         nest = nla_nest_start_noflag(skb, enc_opt_type);
2090         if (!nest)
2091                 goto nla_put_failure;
2092
2093         switch (enc_opts->dst_opt_type) {
2094         case TUNNEL_GENEVE_OPT:
2095                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2096                 if (err)
2097                         goto nla_put_failure;
2098                 break;
2099         default:
2100                 goto nla_put_failure;
2101         }
2102         nla_nest_end(skb, nest);
2103         return 0;
2104
2105 nla_put_failure:
2106         nla_nest_cancel(skb, nest);
2107         return -EMSGSIZE;
2108 }
2109
2110 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2111                                struct flow_dissector_key_enc_opts *key_opts,
2112                                struct flow_dissector_key_enc_opts *msk_opts)
2113 {
2114         int err;
2115
2116         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2117         if (err)
2118                 return err;
2119
2120         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2121 }
2122
2123 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2124                        struct fl_flow_key *key, struct fl_flow_key *mask)
2125 {
2126         if (mask->indev_ifindex) {
2127                 struct net_device *dev;
2128
2129                 dev = __dev_get_by_index(net, key->indev_ifindex);
2130                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2131                         goto nla_put_failure;
2132         }
2133
2134         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2135                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2136                             sizeof(key->eth.dst)) ||
2137             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2138                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2139                             sizeof(key->eth.src)) ||
2140             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2141                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2142                             sizeof(key->basic.n_proto)))
2143                 goto nla_put_failure;
2144
2145         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2146                 goto nla_put_failure;
2147
2148         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2149                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2150                 goto nla_put_failure;
2151
2152         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2153                              TCA_FLOWER_KEY_CVLAN_PRIO,
2154                              &key->cvlan, &mask->cvlan) ||
2155             (mask->cvlan.vlan_tpid &&
2156              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2157                           key->cvlan.vlan_tpid)))
2158                 goto nla_put_failure;
2159
2160         if (mask->basic.n_proto) {
2161                 if (mask->cvlan.vlan_tpid) {
2162                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
2163                                          key->basic.n_proto))
2164                                 goto nla_put_failure;
2165                 } else if (mask->vlan.vlan_tpid) {
2166                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2167                                          key->basic.n_proto))
2168                                 goto nla_put_failure;
2169                 }
2170         }
2171
2172         if ((key->basic.n_proto == htons(ETH_P_IP) ||
2173              key->basic.n_proto == htons(ETH_P_IPV6)) &&
2174             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
2175                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
2176                             sizeof(key->basic.ip_proto)) ||
2177             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
2178                 goto nla_put_failure;
2179
2180         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2181             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
2182                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
2183                              sizeof(key->ipv4.src)) ||
2184              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
2185                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
2186                              sizeof(key->ipv4.dst))))
2187                 goto nla_put_failure;
2188         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2189                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
2190                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
2191                                   sizeof(key->ipv6.src)) ||
2192                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
2193                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
2194                                   sizeof(key->ipv6.dst))))
2195                 goto nla_put_failure;
2196
2197         if (key->basic.ip_proto == IPPROTO_TCP &&
2198             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
2199                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
2200                              sizeof(key->tp.src)) ||
2201              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
2202                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
2203                              sizeof(key->tp.dst)) ||
2204              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
2205                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
2206                              sizeof(key->tcp.flags))))
2207                 goto nla_put_failure;
2208         else if (key->basic.ip_proto == IPPROTO_UDP &&
2209                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
2210                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
2211                                   sizeof(key->tp.src)) ||
2212                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
2213                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
2214                                   sizeof(key->tp.dst))))
2215                 goto nla_put_failure;
2216         else if (key->basic.ip_proto == IPPROTO_SCTP &&
2217                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
2218                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
2219                                   sizeof(key->tp.src)) ||
2220                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
2221                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
2222                                   sizeof(key->tp.dst))))
2223                 goto nla_put_failure;
2224         else if (key->basic.n_proto == htons(ETH_P_IP) &&
2225                  key->basic.ip_proto == IPPROTO_ICMP &&
2226                  (fl_dump_key_val(skb, &key->icmp.type,
2227                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
2228                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
2229                                   sizeof(key->icmp.type)) ||
2230                   fl_dump_key_val(skb, &key->icmp.code,
2231                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
2232                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
2233                                   sizeof(key->icmp.code))))
2234                 goto nla_put_failure;
2235         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
2236                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
2237                  (fl_dump_key_val(skb, &key->icmp.type,
2238                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
2239                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
2240                                   sizeof(key->icmp.type)) ||
2241                   fl_dump_key_val(skb, &key->icmp.code,
2242                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
2243                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
2244                                   sizeof(key->icmp.code))))
2245                 goto nla_put_failure;
2246         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
2247                   key->basic.n_proto == htons(ETH_P_RARP)) &&
2248                  (fl_dump_key_val(skb, &key->arp.sip,
2249                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
2250                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
2251                                   sizeof(key->arp.sip)) ||
2252                   fl_dump_key_val(skb, &key->arp.tip,
2253                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
2254                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
2255                                   sizeof(key->arp.tip)) ||
2256                   fl_dump_key_val(skb, &key->arp.op,
2257                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
2258                                   TCA_FLOWER_KEY_ARP_OP_MASK,
2259                                   sizeof(key->arp.op)) ||
2260                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2261                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2262                                   sizeof(key->arp.sha)) ||
2263                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2264                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2265                                   sizeof(key->arp.tha))))
2266                 goto nla_put_failure;
2267
2268         if ((key->basic.ip_proto == IPPROTO_TCP ||
2269              key->basic.ip_proto == IPPROTO_UDP ||
2270              key->basic.ip_proto == IPPROTO_SCTP) &&
2271              fl_dump_key_port_range(skb, key, mask))
2272                 goto nla_put_failure;
2273
2274         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2275             (fl_dump_key_val(skb, &key->enc_ipv4.src,
2276                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
2277                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
2278                             sizeof(key->enc_ipv4.src)) ||
2279              fl_dump_key_val(skb, &key->enc_ipv4.dst,
2280                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
2281                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
2282                              sizeof(key->enc_ipv4.dst))))
2283                 goto nla_put_failure;
2284         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2285                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
2286                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
2287                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
2288                             sizeof(key->enc_ipv6.src)) ||
2289                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
2290                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
2291                                  &mask->enc_ipv6.dst,
2292                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
2293                             sizeof(key->enc_ipv6.dst))))
2294                 goto nla_put_failure;
2295
2296         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
2297                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
2298                             sizeof(key->enc_key_id)) ||
2299             fl_dump_key_val(skb, &key->enc_tp.src,
2300                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
2301                             &mask->enc_tp.src,
2302                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
2303                             sizeof(key->enc_tp.src)) ||
2304             fl_dump_key_val(skb, &key->enc_tp.dst,
2305                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
2306                             &mask->enc_tp.dst,
2307                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
2308                             sizeof(key->enc_tp.dst)) ||
2309             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
2310             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
2311                 goto nla_put_failure;
2312
2313         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
2314                 goto nla_put_failure;
2315
2316         return 0;
2317
2318 nla_put_failure:
2319         return -EMSGSIZE;
2320 }
2321
2322 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
2323                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
2324 {
2325         struct cls_fl_filter *f = fh;
2326         struct nlattr *nest;
2327         struct fl_flow_key *key, *mask;
2328         bool skip_hw;
2329
2330         if (!f)
2331                 return skb->len;
2332
2333         t->tcm_handle = f->handle;
2334
2335         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2336         if (!nest)
2337                 goto nla_put_failure;
2338
2339         spin_lock(&tp->lock);
2340
2341         if (f->res.classid &&
2342             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
2343                 goto nla_put_failure_locked;
2344
2345         key = &f->key;
2346         mask = &f->mask->key;
2347         skip_hw = tc_skip_hw(f->flags);
2348
2349         if (fl_dump_key(skb, net, key, mask))
2350                 goto nla_put_failure_locked;
2351
2352         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
2353                 goto nla_put_failure_locked;
2354
2355         spin_unlock(&tp->lock);
2356
2357         if (!skip_hw)
2358                 fl_hw_update_stats(tp, f, rtnl_held);
2359
2360         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
2361                 goto nla_put_failure;
2362
2363         if (tcf_exts_dump(skb, &f->exts))
2364                 goto nla_put_failure;
2365
2366         nla_nest_end(skb, nest);
2367
2368         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
2369                 goto nla_put_failure;
2370
2371         return skb->len;
2372
2373 nla_put_failure_locked:
2374         spin_unlock(&tp->lock);
2375 nla_put_failure:
2376         nla_nest_cancel(skb, nest);
2377         return -1;
2378 }
2379
2380 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
2381 {
2382         struct fl_flow_tmplt *tmplt = tmplt_priv;
2383         struct fl_flow_key *key, *mask;
2384         struct nlattr *nest;
2385
2386         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2387         if (!nest)
2388                 goto nla_put_failure;
2389
2390         key = &tmplt->dummy_key;
2391         mask = &tmplt->mask;
2392
2393         if (fl_dump_key(skb, net, key, mask))
2394                 goto nla_put_failure;
2395
2396         nla_nest_end(skb, nest);
2397
2398         return skb->len;
2399
2400 nla_put_failure:
2401         nla_nest_cancel(skb, nest);
2402         return -EMSGSIZE;
2403 }
2404
2405 static void fl_bind_class(void *fh, u32 classid, unsigned long cl)
2406 {
2407         struct cls_fl_filter *f = fh;
2408
2409         if (f && f->res.classid == classid)
2410                 f->res.class = cl;
2411 }
2412
2413 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
2414         .kind           = "flower",
2415         .classify       = fl_classify,
2416         .init           = fl_init,
2417         .destroy        = fl_destroy,
2418         .get            = fl_get,
2419         .put            = fl_put,
2420         .change         = fl_change,
2421         .delete         = fl_delete,
2422         .walk           = fl_walk,
2423         .reoffload      = fl_reoffload,
2424         .dump           = fl_dump,
2425         .bind_class     = fl_bind_class,
2426         .tmplt_create   = fl_tmplt_create,
2427         .tmplt_destroy  = fl_tmplt_destroy,
2428         .tmplt_dump     = fl_tmplt_dump,
2429         .owner          = THIS_MODULE,
2430         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
2431 };
2432
2433 static int __init cls_fl_init(void)
2434 {
2435         return register_tcf_proto_ops(&cls_fl_ops);
2436 }
2437
2438 static void __exit cls_fl_exit(void)
2439 {
2440         unregister_tcf_proto_ops(&cls_fl_ops);
2441 }
2442
2443 module_init(cls_fl_init);
2444 module_exit(cls_fl_exit);
2445
2446 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
2447 MODULE_DESCRIPTION("Flower classifier");
2448 MODULE_LICENSE("GPL v2");