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