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