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                               cls_flower.stats.used_hw_stats,
 497                               cls_flower.stats.used_hw_stats_valid);
 498 }
 499
 500 static void __fl_put(struct cls_fl_filter *f)
 501 {
 502         if (!refcount_dec_and_test(&f->refcnt))
 503                 return;
 504
 505         if (tcf_exts_get_net(&f->exts))
 506                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 507         else
 508                 __fl_destroy_filter(f);
 509 }
 510
 511 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 512 {
 513         struct cls_fl_filter *f;
 514
 515         rcu_read_lock();
 516         f = idr_find(&head->handle_idr, handle);
 517         if (f && !refcount_inc_not_zero(&f->refcnt))
 518                 f = NULL;
 519         rcu_read_unlock();
 520
 521         return f;
 522 }
 523
 524 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 525                        bool *last, bool rtnl_held,
 526                        struct netlink_ext_ack *extack)
 527 {
 528         struct cls_fl_head *head = fl_head_dereference(tp);
 529
 530         *last = false;
 531
 532         spin_lock(&tp->lock);
 533         if (f->deleted) {
 534                 spin_unlock(&tp->lock);
 535                 return -ENOENT;
 536         }
 537
 538         f->deleted = true;
 539         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 540                                f->mask->filter_ht_params);
 541         idr_remove(&head->handle_idr, f->handle);
 542         list_del_rcu(&f->list);
 543         spin_unlock(&tp->lock);
 544
 545         *last = fl_mask_put(head, f->mask);
 546         if (!tc_skip_hw(f->flags))
 547                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 548         tcf_unbind_filter(tp, &f->res);
 549         __fl_put(f);
 550
 551         return 0;
 552 }
 553
 554 static void fl_destroy_sleepable(struct work_struct *work)
 555 {
 556         struct cls_fl_head *head = container_of(to_rcu_work(work),
 557                                                 struct cls_fl_head,
 558                                                 rwork);
 559
 560         rhashtable_destroy(&head->ht);
 561         kfree(head);
 562         module_put(THIS_MODULE);
 563 }
 564
 565 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 566                        struct netlink_ext_ack *extack)
 567 {
 568         struct cls_fl_head *head = fl_head_dereference(tp);
 569         struct fl_flow_mask *mask, *next_mask;
 570         struct cls_fl_filter *f, *next;
 571         bool last;
 572
 573         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 574                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 575                         __fl_delete(tp, f, &last, rtnl_held, extack);
 576                         if (last)
 577                                 break;
 578                 }
 579         }
 580         idr_destroy(&head->handle_idr);
 581
 582         __module_get(THIS_MODULE);
 583         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 584 }
 585
 586 static void fl_put(struct tcf_proto *tp, void *arg)
 587 {
 588         struct cls_fl_filter *f = arg;
 589
 590         __fl_put(f);
 591 }
 592
 593 static void *fl_get(struct tcf_proto *tp, u32 handle)
 594 {
 595         struct cls_fl_head *head = fl_head_dereference(tp);
 596
 597         return __fl_get(head, handle);
 598 }
 599
 600 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 601         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 602         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 603         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 604                                             .len = IFNAMSIZ },
 605         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 606         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 607         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 608         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 609         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 610         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 611         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 612         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 613         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 614         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 615         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 616         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 617         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 618         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 619         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 620         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 621         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 622         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 623         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 624         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 625         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 626         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 627         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 628         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 629         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 630         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 631         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 632         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 633         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 634         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 635         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 636         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 637         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 638         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 639         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 640         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 641         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 642         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 643         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 644         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 645         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 646         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 647         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 648         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 649         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 650         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 651         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 652         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 653         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 654         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 655         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 656         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 657         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 658         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 659         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 660         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 661         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 662         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 663         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 664         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 665         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 666         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 667         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 670         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 671         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 672         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 673         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 674         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 675         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 676         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 677         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 678         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 679         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 680         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 681         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 682         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 683         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 684         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 685         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 686         [TCA_FLOWER_KEY_CT_STATE]       = { .type = NLA_U16 },
 687         [TCA_FLOWER_KEY_CT_STATE_MASK]  = { .type = NLA_U16 },
 688         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 689         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 690         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 691         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 692         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 693                                             .len = 128 / BITS_PER_BYTE },
 694         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 695                                             .len = 128 / BITS_PER_BYTE },
 696         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 697 };
 698
 699 static const struct nla_policy
 700 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 701         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 702                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 703         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 704         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 705         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 706 };
 707
 708 static const struct nla_policy
 709 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 710         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 711         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 712         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 713                                                        .len = 128 },
 714 };
 715
 716 static const struct nla_policy
 717 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 718         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 719 };
 720
 721 static const struct nla_policy
 722 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 723         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 724         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 725         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 726         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 727 };
 728
 729 static void fl_set_key_val(struct nlattr **tb,
 730                            void *val, int val_type,
 731                            void *mask, int mask_type, int len)
 732 {
 733         if (!tb[val_type])
 734                 return;
 735         nla_memcpy(val, tb[val_type], len);
 736         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 737                 memset(mask, 0xff, len);
 738         else
 739                 nla_memcpy(mask, tb[mask_type], len);
 740 }
 741
 742 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 743                                  struct fl_flow_key *mask,
 744                                  struct netlink_ext_ack *extack)
 745 {
 746         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 747                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 748                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 749         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 750                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 751                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 752         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 753                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 754                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 755         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 756                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 757                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 758
 759         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 760             htons(key->tp_range.tp_max.dst) <=
 761             htons(key->tp_range.tp_min.dst)) {
 762                 NL_SET_ERR_MSG_ATTR(extack,
 763                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 764                                     "Invalid destination port range (min must be strictly smaller than max)");
 765                 return -EINVAL;
 766         }
 767         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 768             htons(key->tp_range.tp_max.src) <=
 769             htons(key->tp_range.tp_min.src)) {
 770                 NL_SET_ERR_MSG_ATTR(extack,
 771                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 772                                     "Invalid source port range (min must be strictly smaller than max)");
 773                 return -EINVAL;
 774         }
 775
 776         return 0;
 777 }
 778
 779 static int fl_set_key_mpls(struct nlattr **tb,
 780                            struct flow_dissector_key_mpls *key_val,
 781                            struct flow_dissector_key_mpls *key_mask,
 782                            struct netlink_ext_ack *extack)
 783 {
 784         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 785                 key_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 786                 key_mask->mpls_ttl = MPLS_TTL_MASK;
 787         }
 788         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 789                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 790
 791                 if (bos & ~MPLS_BOS_MASK) {
 792                         NL_SET_ERR_MSG_ATTR(extack,
 793                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
 794                                             "Bottom Of Stack (BOS) must be 0 or 1");
 795                         return -EINVAL;
 796                 }
 797                 key_val->mpls_bos = bos;
 798                 key_mask->mpls_bos = MPLS_BOS_MASK;
 799         }
 800         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 801                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 802
 803                 if (tc & ~MPLS_TC_MASK) {
 804                         NL_SET_ERR_MSG_ATTR(extack,
 805                                             tb[TCA_FLOWER_KEY_MPLS_TC],
 806                                             "Traffic Class (TC) must be between 0 and 7");
 807                         return -EINVAL;
 808                 }
 809                 key_val->mpls_tc = tc;
 810                 key_mask->mpls_tc = MPLS_TC_MASK;
 811         }
 812         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 813                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 814
 815                 if (label & ~MPLS_LABEL_MASK) {
 816                         NL_SET_ERR_MSG_ATTR(extack,
 817                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
 818                                             "Label must be between 0 and 1048575");
 819                         return -EINVAL;
 820                 }
 821                 key_val->mpls_label = label;
 822                 key_mask->mpls_label = MPLS_LABEL_MASK;
 823         }
 824         return 0;
 825 }
 826
 827 static void fl_set_key_vlan(struct nlattr **tb,
 828                             __be16 ethertype,
 829                             int vlan_id_key, int vlan_prio_key,
 830                             struct flow_dissector_key_vlan *key_val,
 831                             struct flow_dissector_key_vlan *key_mask)
 832 {
 833 #define VLAN_PRIORITY_MASK      0x7
 834
 835         if (tb[vlan_id_key]) {
 836                 key_val->vlan_id =
 837                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
 838                 key_mask->vlan_id = VLAN_VID_MASK;
 839         }
 840         if (tb[vlan_prio_key]) {
 841                 key_val->vlan_priority =
 842                         nla_get_u8(tb[vlan_prio_key]) &
 843                         VLAN_PRIORITY_MASK;
 844                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 845         }
 846         key_val->vlan_tpid = ethertype;
 847         key_mask->vlan_tpid = cpu_to_be16(~0);
 848 }
 849
 850 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
 851                             u32 *dissector_key, u32 *dissector_mask,
 852                             u32 flower_flag_bit, u32 dissector_flag_bit)
 853 {
 854         if (flower_mask & flower_flag_bit) {
 855                 *dissector_mask |= dissector_flag_bit;
 856                 if (flower_key & flower_flag_bit)
 857                         *dissector_key |= dissector_flag_bit;
 858         }
 859 }
 860
 861 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
 862                             u32 *flags_mask, struct netlink_ext_ack *extack)
 863 {
 864         u32 key, mask;
 865
 866         /* mask is mandatory for flags */
 867         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
 868                 NL_SET_ERR_MSG(extack, "Missing flags mask");
 869                 return -EINVAL;
 870         }
 871
 872         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
 873         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
 874
 875         *flags_key  = 0;
 876         *flags_mask = 0;
 877
 878         fl_set_key_flag(key, mask, flags_key, flags_mask,
 879                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
 880         fl_set_key_flag(key, mask, flags_key, flags_mask,
 881                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
 882                         FLOW_DIS_FIRST_FRAG);
 883
 884         return 0;
 885 }
 886
 887 static void fl_set_key_ip(struct nlattr **tb, bool encap,
 888                           struct flow_dissector_key_ip *key,
 889                           struct flow_dissector_key_ip *mask)
 890 {
 891         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
 892         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
 893         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
 894         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
 895
 896         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
 897         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
 898 }
 899
 900 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
 901                              int depth, int option_len,
 902                              struct netlink_ext_ack *extack)
 903 {
 904         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
 905         struct nlattr *class = NULL, *type = NULL, *data = NULL;
 906         struct geneve_opt *opt;
 907         int err, data_len = 0;
 908
 909         if (option_len > sizeof(struct geneve_opt))
 910                 data_len = option_len - sizeof(struct geneve_opt);
 911
 912         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
 913         memset(opt, 0xff, option_len);
 914         opt->length = data_len / 4;
 915         opt->r1 = 0;
 916         opt->r2 = 0;
 917         opt->r3 = 0;
 918
 919         /* If no mask has been prodived we assume an exact match. */
 920         if (!depth)
 921                 return sizeof(struct geneve_opt) + data_len;
 922
 923         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
 924                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
 925                 return -EINVAL;
 926         }
 927
 928         err = nla_parse_nested_deprecated(tb,
 929                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
 930                                           nla, geneve_opt_policy, extack);
 931         if (err < 0)
 932                 return err;
 933
 934         /* We are not allowed to omit any of CLASS, TYPE or DATA
 935          * fields from the key.
 936          */
 937         if (!option_len &&
 938             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
 939              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
 940              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
 941                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
 942                 return -EINVAL;
 943         }
 944
 945         /* Omitting any of CLASS, TYPE or DATA fields is allowed
 946          * for the mask.
 947          */
 948         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
 949                 int new_len = key->enc_opts.len;
 950
 951                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
 952                 data_len = nla_len(data);
 953                 if (data_len < 4) {
 954                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
 955                         return -ERANGE;
 956                 }
 957                 if (data_len % 4) {
 958                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
 959                         return -ERANGE;
 960                 }
 961
 962                 new_len += sizeof(struct geneve_opt) + data_len;
 963                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
 964                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
 965                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
 966                         return -ERANGE;
 967                 }
 968                 opt->length = data_len / 4;
 969                 memcpy(opt->opt_data, nla_data(data), data_len);
 970         }
 971
 972         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
 973                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
 974                 opt->opt_class = nla_get_be16(class);
 975         }
 976
 977         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
 978                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
 979                 opt->type = nla_get_u8(type);
 980         }
 981
 982         return sizeof(struct geneve_opt) + data_len;
 983 }
 984
 985 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
 986                             int depth, int option_len,
 987                             struct netlink_ext_ack *extack)
 988 {
 989         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
 990         struct vxlan_metadata *md;
 991         int err;
 992
 993         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
 994         memset(md, 0xff, sizeof(*md));
 995
 996         if (!depth)
 997                 return sizeof(*md);
 998
 999         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1000                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1001                 return -EINVAL;
1002         }
1003
1004         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1005                                vxlan_opt_policy, extack);
1006         if (err < 0)
1007                 return err;
1008
1009         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1010                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1011                 return -EINVAL;
1012         }
1013
1014         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP])
1015                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1016
1017         return sizeof(*md);
1018 }
1019
1020 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1021                              int depth, int option_len,
1022                              struct netlink_ext_ack *extack)
1023 {
1024         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1025         struct erspan_metadata *md;
1026         int err;
1027
1028         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1029         memset(md, 0xff, sizeof(*md));
1030         md->version = 1;
1031
1032         if (!depth)
1033                 return sizeof(*md);
1034
1035         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1036                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1037                 return -EINVAL;
1038         }
1039
1040         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1041                                erspan_opt_policy, extack);
1042         if (err < 0)
1043                 return err;
1044
1045         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1046                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1047                 return -EINVAL;
1048         }
1049
1050         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1051                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1052
1053         if (md->version == 1) {
1054                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1055                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1056                         return -EINVAL;
1057                 }
1058                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1059                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1060                         md->u.index = nla_get_be32(nla);
1061                 }
1062         } else if (md->version == 2) {
1063                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1064                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1065                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1066                         return -EINVAL;
1067                 }
1068                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1069                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1070                         md->u.md2.dir = nla_get_u8(nla);
1071                 }
1072                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1073                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1074                         set_hwid(&md->u.md2, nla_get_u8(nla));
1075                 }
1076         } else {
1077                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1078                 return -EINVAL;
1079         }
1080
1081         return sizeof(*md);
1082 }
1083
1084 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1085                           struct fl_flow_key *mask,
1086                           struct netlink_ext_ack *extack)
1087 {
1088         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1089         int err, option_len, key_depth, msk_depth = 0;
1090
1091         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1092                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1093                                              enc_opts_policy, extack);
1094         if (err)
1095                 return err;
1096
1097         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1098
1099         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1100                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1101                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1102                                                      enc_opts_policy, extack);
1103                 if (err)
1104                         return err;
1105
1106                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1107                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1108         }
1109
1110         nla_for_each_attr(nla_opt_key, nla_enc_key,
1111                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1112                 switch (nla_type(nla_opt_key)) {
1113                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1114                         if (key->enc_opts.dst_opt_type &&
1115                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1116                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1117                                 return -EINVAL;
1118                         }
1119                         option_len = 0;
1120                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1121                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1122                                                        key_depth, option_len,
1123                                                        extack);
1124                         if (option_len < 0)
1125                                 return option_len;
1126
1127                         key->enc_opts.len += option_len;
1128                         /* At the same time we need to parse through the mask
1129                          * in order to verify exact and mask attribute lengths.
1130                          */
1131                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1132                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1133                                                        msk_depth, option_len,
1134                                                        extack);
1135                         if (option_len < 0)
1136                                 return option_len;
1137
1138                         mask->enc_opts.len += option_len;
1139                         if (key->enc_opts.len != mask->enc_opts.len) {
1140                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1141                                 return -EINVAL;
1142                         }
1143
1144                         if (msk_depth)
1145                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1146                         break;
1147                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1148                         if (key->enc_opts.dst_opt_type) {
1149                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1150                                 return -EINVAL;
1151                         }
1152                         option_len = 0;
1153                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1154                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1155                                                       key_depth, option_len,
1156                                                       extack);
1157                         if (option_len < 0)
1158                                 return option_len;
1159
1160                         key->enc_opts.len += option_len;
1161                         /* At the same time we need to parse through the mask
1162                          * in order to verify exact and mask attribute lengths.
1163                          */
1164                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1165                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1166                                                       msk_depth, option_len,
1167                                                       extack);
1168                         if (option_len < 0)
1169                                 return option_len;
1170
1171                         mask->enc_opts.len += option_len;
1172                         if (key->enc_opts.len != mask->enc_opts.len) {
1173                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1174                                 return -EINVAL;
1175                         }
1176
1177                         if (msk_depth)
1178                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1179                         break;
1180                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1181                         if (key->enc_opts.dst_opt_type) {
1182                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1183                                 return -EINVAL;
1184                         }
1185                         option_len = 0;
1186                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1187                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1188                                                        key_depth, option_len,
1189                                                        extack);
1190                         if (option_len < 0)
1191                                 return option_len;
1192
1193                         key->enc_opts.len += option_len;
1194                         /* At the same time we need to parse through the mask
1195                          * in order to verify exact and mask attribute lengths.
1196                          */
1197                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1198                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1199                                                        msk_depth, option_len,
1200                                                        extack);
1201                         if (option_len < 0)
1202                                 return option_len;
1203
1204                         mask->enc_opts.len += option_len;
1205                         if (key->enc_opts.len != mask->enc_opts.len) {
1206                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1207                                 return -EINVAL;
1208                         }
1209
1210                         if (msk_depth)
1211                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1212                         break;
1213                 default:
1214                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1215                         return -EINVAL;
1216                 }
1217         }
1218
1219         return 0;
1220 }
1221
1222 static int fl_set_key_ct(struct nlattr **tb,
1223                          struct flow_dissector_key_ct *key,
1224                          struct flow_dissector_key_ct *mask,
1225                          struct netlink_ext_ack *extack)
1226 {
1227         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1228                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1229                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1230                         return -EOPNOTSUPP;
1231                 }
1232                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1233                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1234                                sizeof(key->ct_state));
1235         }
1236         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1237                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1238                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1239                         return -EOPNOTSUPP;
1240                 }
1241                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1242                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1243                                sizeof(key->ct_zone));
1244         }
1245         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1246                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1247                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1248                         return -EOPNOTSUPP;
1249                 }
1250                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1251                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1252                                sizeof(key->ct_mark));
1253         }
1254         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1255                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1256                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1257                         return -EOPNOTSUPP;
1258                 }
1259                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1260                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1261                                sizeof(key->ct_labels));
1262         }
1263
1264         return 0;
1265 }
1266
1267 static int fl_set_key(struct net *net, struct nlattr **tb,
1268                       struct fl_flow_key *key, struct fl_flow_key *mask,
1269                       struct netlink_ext_ack *extack)
1270 {
1271         __be16 ethertype;
1272         int ret = 0;
1273
1274         if (tb[TCA_FLOWER_INDEV]) {
1275                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1276                 if (err < 0)
1277                         return err;
1278                 key->meta.ingress_ifindex = err;
1279                 mask->meta.ingress_ifindex = 0xffffffff;
1280         }
1281
1282         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1283                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1284                        sizeof(key->eth.dst));
1285         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1286                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1287                        sizeof(key->eth.src));
1288
1289         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1290                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1291
1292                 if (eth_type_vlan(ethertype)) {
1293                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1294                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
1295                                         &mask->vlan);
1296
1297                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1298                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1299                                 if (eth_type_vlan(ethertype)) {
1300                                         fl_set_key_vlan(tb, ethertype,
1301                                                         TCA_FLOWER_KEY_CVLAN_ID,
1302                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1303                                                         &key->cvlan, &mask->cvlan);
1304                                         fl_set_key_val(tb, &key->basic.n_proto,
1305                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1306                                                        &mask->basic.n_proto,
1307                                                        TCA_FLOWER_UNSPEC,
1308                                                        sizeof(key->basic.n_proto));
1309                                 } else {
1310                                         key->basic.n_proto = ethertype;
1311                                         mask->basic.n_proto = cpu_to_be16(~0);
1312                                 }
1313                         }
1314                 } else {
1315                         key->basic.n_proto = ethertype;
1316                         mask->basic.n_proto = cpu_to_be16(~0);
1317                 }
1318         }
1319
1320         if (key->basic.n_proto == htons(ETH_P_IP) ||
1321             key->basic.n_proto == htons(ETH_P_IPV6)) {
1322                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1323                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1324                                sizeof(key->basic.ip_proto));
1325                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1326         }
1327
1328         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1329                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1330                 mask->control.addr_type = ~0;
1331                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1332                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1333                                sizeof(key->ipv4.src));
1334                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1335                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1336                                sizeof(key->ipv4.dst));
1337         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1338                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1339                 mask->control.addr_type = ~0;
1340                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1341                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1342                                sizeof(key->ipv6.src));
1343                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1344                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1345                                sizeof(key->ipv6.dst));
1346         }
1347
1348         if (key->basic.ip_proto == IPPROTO_TCP) {
1349                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1350                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1351                                sizeof(key->tp.src));
1352                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1353                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1354                                sizeof(key->tp.dst));
1355                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1356                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1357                                sizeof(key->tcp.flags));
1358         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1359                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1360                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1361                                sizeof(key->tp.src));
1362                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1363                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1364                                sizeof(key->tp.dst));
1365         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1366                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1367                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1368                                sizeof(key->tp.src));
1369                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1370                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1371                                sizeof(key->tp.dst));
1372         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1373                    key->basic.ip_proto == IPPROTO_ICMP) {
1374                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1375                                &mask->icmp.type,
1376                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1377                                sizeof(key->icmp.type));
1378                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1379                                &mask->icmp.code,
1380                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1381                                sizeof(key->icmp.code));
1382         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1383                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1384                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1385                                &mask->icmp.type,
1386                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1387                                sizeof(key->icmp.type));
1388                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1389                                &mask->icmp.code,
1390                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1391                                sizeof(key->icmp.code));
1392         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1393                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1394                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1395                 if (ret)
1396                         return ret;
1397         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1398                    key->basic.n_proto == htons(ETH_P_RARP)) {
1399                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1400                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1401                                sizeof(key->arp.sip));
1402                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1403                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1404                                sizeof(key->arp.tip));
1405                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1406                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1407                                sizeof(key->arp.op));
1408                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1409                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1410                                sizeof(key->arp.sha));
1411                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1412                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1413                                sizeof(key->arp.tha));
1414         }
1415
1416         if (key->basic.ip_proto == IPPROTO_TCP ||
1417             key->basic.ip_proto == IPPROTO_UDP ||
1418             key->basic.ip_proto == IPPROTO_SCTP) {
1419                 ret = fl_set_key_port_range(tb, key, mask, extack);
1420                 if (ret)
1421                         return ret;
1422         }
1423
1424         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1425             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1426                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1427                 mask->enc_control.addr_type = ~0;
1428                 fl_set_key_val(tb, &key->enc_ipv4.src,
1429                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1430                                &mask->enc_ipv4.src,
1431                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1432                                sizeof(key->enc_ipv4.src));
1433                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1434                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1435                                &mask->enc_ipv4.dst,
1436                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1437                                sizeof(key->enc_ipv4.dst));
1438         }
1439
1440         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1441             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1442                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1443                 mask->enc_control.addr_type = ~0;
1444                 fl_set_key_val(tb, &key->enc_ipv6.src,
1445                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1446                                &mask->enc_ipv6.src,
1447                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1448                                sizeof(key->enc_ipv6.src));
1449                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1450                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1451                                &mask->enc_ipv6.dst,
1452                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1453                                sizeof(key->enc_ipv6.dst));
1454         }
1455
1456         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1457                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1458                        sizeof(key->enc_key_id.keyid));
1459
1460         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1461                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1462                        sizeof(key->enc_tp.src));
1463
1464         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1465                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1466                        sizeof(key->enc_tp.dst));
1467
1468         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1469
1470         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1471                 ret = fl_set_enc_opt(tb, key, mask, extack);
1472                 if (ret)
1473                         return ret;
1474         }
1475
1476         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1477         if (ret)
1478                 return ret;
1479
1480         if (tb[TCA_FLOWER_KEY_FLAGS])
1481                 ret = fl_set_key_flags(tb, &key->control.flags,
1482                                        &mask->control.flags, extack);
1483
1484         return ret;
1485 }
1486
1487 static void fl_mask_copy(struct fl_flow_mask *dst,
1488                          struct fl_flow_mask *src)
1489 {
1490         const void *psrc = fl_key_get_start(&src->key, src);
1491         void *pdst = fl_key_get_start(&dst->key, src);
1492
1493         memcpy(pdst, psrc, fl_mask_range(src));
1494         dst->range = src->range;
1495 }
1496
1497 static const struct rhashtable_params fl_ht_params = {
1498         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1499         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1500         .automatic_shrinking = true,
1501 };
1502
1503 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1504 {
1505         mask->filter_ht_params = fl_ht_params;
1506         mask->filter_ht_params.key_len = fl_mask_range(mask);
1507         mask->filter_ht_params.key_offset += mask->range.start;
1508
1509         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1510 }
1511
1512 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1513 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1514
1515 #define FL_KEY_IS_MASKED(mask, member)                                          \
1516         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1517                    0, FL_KEY_MEMBER_SIZE(member))                               \
1518
1519 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1520         do {                                                                    \
1521                 keys[cnt].key_id = id;                                          \
1522                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1523                 cnt++;                                                          \
1524         } while(0);
1525
1526 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1527         do {                                                                    \
1528                 if (FL_KEY_IS_MASKED(mask, member))                             \
1529                         FL_KEY_SET(keys, cnt, id, member);                      \
1530         } while(0);
1531
1532 static void fl_init_dissector(struct flow_dissector *dissector,
1533                               struct fl_flow_key *mask)
1534 {
1535         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1536         size_t cnt = 0;
1537
1538         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1539                              FLOW_DISSECTOR_KEY_META, meta);
1540         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1541         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1542         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1543                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1544         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1545                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1546         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1547                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1548         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1549                              FLOW_DISSECTOR_KEY_PORTS, tp);
1550         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1551                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1552         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1553                              FLOW_DISSECTOR_KEY_IP, ip);
1554         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1555                              FLOW_DISSECTOR_KEY_TCP, tcp);
1556         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1557                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1558         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1559                              FLOW_DISSECTOR_KEY_ARP, arp);
1560         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1561                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1562         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1563                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1564         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1565                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1566         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1567                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1568         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1569                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1570         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1571                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1572         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1573             FL_KEY_IS_MASKED(mask, enc_ipv6))
1574                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1575                            enc_control);
1576         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1577                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1578         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1579                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1580         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1581                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1582         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1583                              FLOW_DISSECTOR_KEY_CT, ct);
1584
1585         skb_flow_dissector_init(dissector, keys, cnt);
1586 }
1587
1588 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1589                                                struct fl_flow_mask *mask)
1590 {
1591         struct fl_flow_mask *newmask;
1592         int err;
1593
1594         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1595         if (!newmask)
1596                 return ERR_PTR(-ENOMEM);
1597
1598         fl_mask_copy(newmask, mask);
1599
1600         if ((newmask->key.tp_range.tp_min.dst &&
1601              newmask->key.tp_range.tp_max.dst) ||
1602             (newmask->key.tp_range.tp_min.src &&
1603              newmask->key.tp_range.tp_max.src))
1604                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1605
1606         err = fl_init_mask_hashtable(newmask);
1607         if (err)
1608                 goto errout_free;
1609
1610         fl_init_dissector(&newmask->dissector, &newmask->key);
1611
1612         INIT_LIST_HEAD_RCU(&newmask->filters);
1613
1614         refcount_set(&newmask->refcnt, 1);
1615         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1616                                       &newmask->ht_node, mask_ht_params);
1617         if (err)
1618                 goto errout_destroy;
1619
1620         spin_lock(&head->masks_lock);
1621         list_add_tail_rcu(&newmask->list, &head->masks);
1622         spin_unlock(&head->masks_lock);
1623
1624         return newmask;
1625
1626 errout_destroy:
1627         rhashtable_destroy(&newmask->ht);
1628 errout_free:
1629         kfree(newmask);
1630
1631         return ERR_PTR(err);
1632 }
1633
1634 static int fl_check_assign_mask(struct cls_fl_head *head,
1635                                 struct cls_fl_filter *fnew,
1636                                 struct cls_fl_filter *fold,
1637                                 struct fl_flow_mask *mask)
1638 {
1639         struct fl_flow_mask *newmask;
1640         int ret = 0;
1641
1642         rcu_read_lock();
1643
1644         /* Insert mask as temporary node to prevent concurrent creation of mask
1645          * with same key. Any concurrent lookups with same key will return
1646          * -EAGAIN because mask's refcnt is zero.
1647          */
1648         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1649                                                        &mask->ht_node,
1650                                                        mask_ht_params);
1651         if (!fnew->mask) {
1652                 rcu_read_unlock();
1653
1654                 if (fold) {
1655                         ret = -EINVAL;
1656                         goto errout_cleanup;
1657                 }
1658
1659                 newmask = fl_create_new_mask(head, mask);
1660                 if (IS_ERR(newmask)) {
1661                         ret = PTR_ERR(newmask);
1662                         goto errout_cleanup;
1663                 }
1664
1665                 fnew->mask = newmask;
1666                 return 0;
1667         } else if (IS_ERR(fnew->mask)) {
1668                 ret = PTR_ERR(fnew->mask);
1669         } else if (fold && fold->mask != fnew->mask) {
1670                 ret = -EINVAL;
1671         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1672                 /* Mask was deleted concurrently, try again */
1673                 ret = -EAGAIN;
1674         }
1675         rcu_read_unlock();
1676         return ret;
1677
1678 errout_cleanup:
1679         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1680                                mask_ht_params);
1681         return ret;
1682 }
1683
1684 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1685                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1686                         unsigned long base, struct nlattr **tb,
1687                         struct nlattr *est, bool ovr,
1688                         struct fl_flow_tmplt *tmplt, bool rtnl_held,
1689                         struct netlink_ext_ack *extack)
1690 {
1691         int err;
1692
1693         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, rtnl_held,
1694                                 extack);
1695         if (err < 0)
1696                 return err;
1697
1698         if (tb[TCA_FLOWER_CLASSID]) {
1699                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1700                 if (!rtnl_held)
1701                         rtnl_lock();
1702                 tcf_bind_filter(tp, &f->res, base);
1703                 if (!rtnl_held)
1704                         rtnl_unlock();
1705         }
1706
1707         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1708         if (err)
1709                 return err;
1710
1711         fl_mask_update_range(mask);
1712         fl_set_masked_key(&f->mkey, &f->key, mask);
1713
1714         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1715                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1716                 return -EINVAL;
1717         }
1718
1719         return 0;
1720 }
1721
1722 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
1723                                struct cls_fl_filter *fold,
1724                                bool *in_ht)
1725 {
1726         struct fl_flow_mask *mask = fnew->mask;
1727         int err;
1728
1729         err = rhashtable_lookup_insert_fast(&mask->ht,
1730                                             &fnew->ht_node,
1731                                             mask->filter_ht_params);
1732         if (err) {
1733                 *in_ht = false;
1734                 /* It is okay if filter with same key exists when
1735                  * overwriting.
1736                  */
1737                 return fold && err == -EEXIST ? 0 : err;
1738         }
1739
1740         *in_ht = true;
1741         return 0;
1742 }
1743
1744 static int fl_change(struct net *net, struct sk_buff *in_skb,
1745                      struct tcf_proto *tp, unsigned long base,
1746                      u32 handle, struct nlattr **tca,
1747                      void **arg, bool ovr, bool rtnl_held,
1748                      struct netlink_ext_ack *extack)
1749 {
1750         struct cls_fl_head *head = fl_head_dereference(tp);
1751         struct cls_fl_filter *fold = *arg;
1752         struct cls_fl_filter *fnew;
1753         struct fl_flow_mask *mask;
1754         struct nlattr **tb;
1755         bool in_ht;
1756         int err;
1757
1758         if (!tca[TCA_OPTIONS]) {
1759                 err = -EINVAL;
1760                 goto errout_fold;
1761         }
1762
1763         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
1764         if (!mask) {
1765                 err = -ENOBUFS;
1766                 goto errout_fold;
1767         }
1768
1769         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1770         if (!tb) {
1771                 err = -ENOBUFS;
1772                 goto errout_mask_alloc;
1773         }
1774
1775         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1776                                           tca[TCA_OPTIONS], fl_policy, NULL);
1777         if (err < 0)
1778                 goto errout_tb;
1779
1780         if (fold && handle && fold->handle != handle) {
1781                 err = -EINVAL;
1782                 goto errout_tb;
1783         }
1784
1785         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
1786         if (!fnew) {
1787                 err = -ENOBUFS;
1788                 goto errout_tb;
1789         }
1790         INIT_LIST_HEAD(&fnew->hw_list);
1791         refcount_set(&fnew->refcnt, 1);
1792
1793         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
1794         if (err < 0)
1795                 goto errout;
1796
1797         if (tb[TCA_FLOWER_FLAGS]) {
1798                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
1799
1800                 if (!tc_flags_valid(fnew->flags)) {
1801                         err = -EINVAL;
1802                         goto errout;
1803                 }
1804         }
1805
1806         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE], ovr,
1807                            tp->chain->tmplt_priv, rtnl_held, extack);
1808         if (err)
1809                 goto errout;
1810
1811         err = fl_check_assign_mask(head, fnew, fold, mask);
1812         if (err)
1813                 goto errout;
1814
1815         err = fl_ht_insert_unique(fnew, fold, &in_ht);
1816         if (err)
1817                 goto errout_mask;
1818
1819         if (!tc_skip_hw(fnew->flags)) {
1820                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
1821                 if (err)
1822                         goto errout_ht;
1823         }
1824
1825         if (!tc_in_hw(fnew->flags))
1826                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1827
1828         spin_lock(&tp->lock);
1829
1830         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
1831          * proto again or create new one, if necessary.
1832          */
1833         if (tp->deleting) {
1834                 err = -EAGAIN;
1835                 goto errout_hw;
1836         }
1837
1838         if (fold) {
1839                 /* Fold filter was deleted concurrently. Retry lookup. */
1840                 if (fold->deleted) {
1841                         err = -EAGAIN;
1842                         goto errout_hw;
1843                 }
1844
1845                 fnew->handle = handle;
1846
1847                 if (!in_ht) {
1848                         struct rhashtable_params params =
1849                                 fnew->mask->filter_ht_params;
1850
1851                         err = rhashtable_insert_fast(&fnew->mask->ht,
1852                                                      &fnew->ht_node,
1853                                                      params);
1854                         if (err)
1855                                 goto errout_hw;
1856                         in_ht = true;
1857                 }
1858
1859                 refcount_inc(&fnew->refcnt);
1860                 rhashtable_remove_fast(&fold->mask->ht,
1861                                        &fold->ht_node,
1862                                        fold->mask->filter_ht_params);
1863                 idr_replace(&head->handle_idr, fnew, fnew->handle);
1864                 list_replace_rcu(&fold->list, &fnew->list);
1865                 fold->deleted = true;
1866
1867                 spin_unlock(&tp->lock);
1868
1869                 fl_mask_put(head, fold->mask);
1870                 if (!tc_skip_hw(fold->flags))
1871                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
1872                 tcf_unbind_filter(tp, &fold->res);
1873                 /* Caller holds reference to fold, so refcnt is always > 0
1874                  * after this.
1875                  */
1876                 refcount_dec(&fold->refcnt);
1877                 __fl_put(fold);
1878         } else {
1879                 if (handle) {
1880                         /* user specifies a handle and it doesn't exist */
1881                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1882                                             handle, GFP_ATOMIC);
1883
1884                         /* Filter with specified handle was concurrently
1885                          * inserted after initial check in cls_api. This is not
1886                          * necessarily an error if NLM_F_EXCL is not set in
1887                          * message flags. Returning EAGAIN will cause cls_api to
1888                          * try to update concurrently inserted rule.
1889                          */
1890                         if (err == -ENOSPC)
1891                                 err = -EAGAIN;
1892                 } else {
1893                         handle = 1;
1894                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1895                                             INT_MAX, GFP_ATOMIC);
1896                 }
1897                 if (err)
1898                         goto errout_hw;
1899
1900                 refcount_inc(&fnew->refcnt);
1901                 fnew->handle = handle;
1902                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
1903                 spin_unlock(&tp->lock);
1904         }
1905
1906         *arg = fnew;
1907
1908         kfree(tb);
1909         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
1910         return 0;
1911
1912 errout_ht:
1913         spin_lock(&tp->lock);
1914 errout_hw:
1915         fnew->deleted = true;
1916         spin_unlock(&tp->lock);
1917         if (!tc_skip_hw(fnew->flags))
1918                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
1919         if (in_ht)
1920                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
1921                                        fnew->mask->filter_ht_params);
1922 errout_mask:
1923         fl_mask_put(head, fnew->mask);
1924 errout:
1925         __fl_put(fnew);
1926 errout_tb:
1927         kfree(tb);
1928 errout_mask_alloc:
1929         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
1930 errout_fold:
1931         if (fold)
1932                 __fl_put(fold);
1933         return err;
1934 }
1935
1936 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
1937                      bool rtnl_held, struct netlink_ext_ack *extack)
1938 {
1939         struct cls_fl_head *head = fl_head_dereference(tp);
1940         struct cls_fl_filter *f = arg;
1941         bool last_on_mask;
1942         int err = 0;
1943
1944         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
1945         *last = list_empty(&head->masks);
1946         __fl_put(f);
1947
1948         return err;
1949 }
1950
1951 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
1952                     bool rtnl_held)
1953 {
1954         struct cls_fl_head *head = fl_head_dereference(tp);
1955         unsigned long id = arg->cookie, tmp;
1956         struct cls_fl_filter *f;
1957
1958         arg->count = arg->skip;
1959
1960         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
1961                 /* don't return filters that are being deleted */
1962                 if (!refcount_inc_not_zero(&f->refcnt))
1963                         continue;
1964                 if (arg->fn(tp, f, arg) < 0) {
1965                         __fl_put(f);
1966                         arg->stop = 1;
1967                         break;
1968                 }
1969                 __fl_put(f);
1970                 arg->count++;
1971         }
1972         arg->cookie = id;
1973 }
1974
1975 static struct cls_fl_filter *
1976 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
1977 {
1978         struct cls_fl_head *head = fl_head_dereference(tp);
1979
1980         spin_lock(&tp->lock);
1981         if (list_empty(&head->hw_filters)) {
1982                 spin_unlock(&tp->lock);
1983                 return NULL;
1984         }
1985
1986         if (!f)
1987                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
1988                                hw_list);
1989         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
1990                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
1991                         spin_unlock(&tp->lock);
1992                         return f;
1993                 }
1994         }
1995
1996         spin_unlock(&tp->lock);
1997         return NULL;
1998 }
1999
2000 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2001                         void *cb_priv, struct netlink_ext_ack *extack)
2002 {
2003         struct tcf_block *block = tp->chain->block;
2004         struct flow_cls_offload cls_flower = {};
2005         struct cls_fl_filter *f = NULL;
2006         int err;
2007
2008         /* hw_filters list can only be changed by hw offload functions after
2009          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2010          * iterating it.
2011          */
2012         ASSERT_RTNL();
2013
2014         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2015                 cls_flower.rule =
2016                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2017                 if (!cls_flower.rule) {
2018                         __fl_put(f);
2019                         return -ENOMEM;
2020                 }
2021
2022                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2023                                            extack);
2024                 cls_flower.command = add ?
2025                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2026                 cls_flower.cookie = (unsigned long)f;
2027                 cls_flower.rule->match.dissector = &f->mask->dissector;
2028                 cls_flower.rule->match.mask = &f->mask->key;
2029                 cls_flower.rule->match.key = &f->mkey;
2030
2031                 err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
2032                 if (err) {
2033                         kfree(cls_flower.rule);
2034                         if (tc_skip_sw(f->flags)) {
2035                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
2036                                 __fl_put(f);
2037                                 return err;
2038                         }
2039                         goto next_flow;
2040                 }
2041
2042                 cls_flower.classid = f->res.classid;
2043
2044                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2045                                             TC_SETUP_CLSFLOWER, &cls_flower,
2046                                             cb_priv, &f->flags,
2047                                             &f->in_hw_count);
2048                 tc_cleanup_flow_action(&cls_flower.rule->action);
2049                 kfree(cls_flower.rule);
2050
2051                 if (err) {
2052                         __fl_put(f);
2053                         return err;
2054                 }
2055 next_flow:
2056                 __fl_put(f);
2057         }
2058
2059         return 0;
2060 }
2061
2062 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2063 {
2064         struct flow_cls_offload *cls_flower = type_data;
2065         struct cls_fl_filter *f =
2066                 (struct cls_fl_filter *) cls_flower->cookie;
2067         struct cls_fl_head *head = fl_head_dereference(tp);
2068
2069         spin_lock(&tp->lock);
2070         list_add(&f->hw_list, &head->hw_filters);
2071         spin_unlock(&tp->lock);
2072 }
2073
2074 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2075 {
2076         struct flow_cls_offload *cls_flower = type_data;
2077         struct cls_fl_filter *f =
2078                 (struct cls_fl_filter *) cls_flower->cookie;
2079
2080         spin_lock(&tp->lock);
2081         if (!list_empty(&f->hw_list))
2082                 list_del_init(&f->hw_list);
2083         spin_unlock(&tp->lock);
2084 }
2085
2086 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2087                               struct fl_flow_tmplt *tmplt)
2088 {
2089         struct flow_cls_offload cls_flower = {};
2090         struct tcf_block *block = chain->block;
2091
2092         cls_flower.rule = flow_rule_alloc(0);
2093         if (!cls_flower.rule)
2094                 return -ENOMEM;
2095
2096         cls_flower.common.chain_index = chain->index;
2097         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2098         cls_flower.cookie = (unsigned long) tmplt;
2099         cls_flower.rule->match.dissector = &tmplt->dissector;
2100         cls_flower.rule->match.mask = &tmplt->mask;
2101         cls_flower.rule->match.key = &tmplt->dummy_key;
2102
2103         /* We don't care if driver (any of them) fails to handle this
2104          * call. It serves just as a hint for it.
2105          */
2106         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2107         kfree(cls_flower.rule);
2108
2109         return 0;
2110 }
2111
2112 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2113                                 struct fl_flow_tmplt *tmplt)
2114 {
2115         struct flow_cls_offload cls_flower = {};
2116         struct tcf_block *block = chain->block;
2117
2118         cls_flower.common.chain_index = chain->index;
2119         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2120         cls_flower.cookie = (unsigned long) tmplt;
2121
2122         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2123 }
2124
2125 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2126                              struct nlattr **tca,
2127                              struct netlink_ext_ack *extack)
2128 {
2129         struct fl_flow_tmplt *tmplt;
2130         struct nlattr **tb;
2131         int err;
2132
2133         if (!tca[TCA_OPTIONS])
2134                 return ERR_PTR(-EINVAL);
2135
2136         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2137         if (!tb)
2138                 return ERR_PTR(-ENOBUFS);
2139         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2140                                           tca[TCA_OPTIONS], fl_policy, NULL);
2141         if (err)
2142                 goto errout_tb;
2143
2144         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2145         if (!tmplt) {
2146                 err = -ENOMEM;
2147                 goto errout_tb;
2148         }
2149         tmplt->chain = chain;
2150         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2151         if (err)
2152                 goto errout_tmplt;
2153
2154         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2155
2156         err = fl_hw_create_tmplt(chain, tmplt);
2157         if (err)
2158                 goto errout_tmplt;
2159
2160         kfree(tb);
2161         return tmplt;
2162
2163 errout_tmplt:
2164         kfree(tmplt);
2165 errout_tb:
2166         kfree(tb);
2167         return ERR_PTR(err);
2168 }
2169
2170 static void fl_tmplt_destroy(void *tmplt_priv)
2171 {
2172         struct fl_flow_tmplt *tmplt = tmplt_priv;
2173
2174         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2175         kfree(tmplt);
2176 }
2177
2178 static int fl_dump_key_val(struct sk_buff *skb,
2179                            void *val, int val_type,
2180                            void *mask, int mask_type, int len)
2181 {
2182         int err;
2183
2184         if (!memchr_inv(mask, 0, len))
2185                 return 0;
2186         err = nla_put(skb, val_type, len, val);
2187         if (err)
2188                 return err;
2189         if (mask_type != TCA_FLOWER_UNSPEC) {
2190                 err = nla_put(skb, mask_type, len, mask);
2191                 if (err)
2192                         return err;
2193         }
2194         return 0;
2195 }
2196
2197 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2198                                   struct fl_flow_key *mask)
2199 {
2200         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2201                             TCA_FLOWER_KEY_PORT_DST_MIN,
2202                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2203                             sizeof(key->tp_range.tp_min.dst)) ||
2204             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2205                             TCA_FLOWER_KEY_PORT_DST_MAX,
2206                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2207                             sizeof(key->tp_range.tp_max.dst)) ||
2208             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2209                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2210                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2211                             sizeof(key->tp_range.tp_min.src)) ||
2212             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2213                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2214                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2215                             sizeof(key->tp_range.tp_max.src)))
2216                 return -1;
2217
2218         return 0;
2219 }
2220
2221 static int fl_dump_key_mpls(struct sk_buff *skb,
2222                             struct flow_dissector_key_mpls *mpls_key,
2223                             struct flow_dissector_key_mpls *mpls_mask)
2224 {
2225         int err;
2226
2227         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
2228                 return 0;
2229         if (mpls_mask->mpls_ttl) {
2230                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2231                                  mpls_key->mpls_ttl);
2232                 if (err)
2233                         return err;
2234         }
2235         if (mpls_mask->mpls_tc) {
2236                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2237                                  mpls_key->mpls_tc);
2238                 if (err)
2239                         return err;
2240         }
2241         if (mpls_mask->mpls_label) {
2242                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2243                                   mpls_key->mpls_label);
2244                 if (err)
2245                         return err;
2246         }
2247         if (mpls_mask->mpls_bos) {
2248                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2249                                  mpls_key->mpls_bos);
2250                 if (err)
2251                         return err;
2252         }
2253         return 0;
2254 }
2255
2256 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2257                           struct flow_dissector_key_ip *key,
2258                           struct flow_dissector_key_ip *mask)
2259 {
2260         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2261         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2262         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2263         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2264
2265         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2266             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2267                 return -1;
2268
2269         return 0;
2270 }
2271
2272 static int fl_dump_key_vlan(struct sk_buff *skb,
2273                             int vlan_id_key, int vlan_prio_key,
2274                             struct flow_dissector_key_vlan *vlan_key,
2275                             struct flow_dissector_key_vlan *vlan_mask)
2276 {
2277         int err;
2278
2279         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2280                 return 0;
2281         if (vlan_mask->vlan_id) {
2282                 err = nla_put_u16(skb, vlan_id_key,
2283                                   vlan_key->vlan_id);
2284                 if (err)
2285                         return err;
2286         }
2287         if (vlan_mask->vlan_priority) {
2288                 err = nla_put_u8(skb, vlan_prio_key,
2289                                  vlan_key->vlan_priority);
2290                 if (err)
2291                         return err;
2292         }
2293         return 0;
2294 }
2295
2296 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2297                             u32 *flower_key, u32 *flower_mask,
2298                             u32 flower_flag_bit, u32 dissector_flag_bit)
2299 {
2300         if (dissector_mask & dissector_flag_bit) {
2301                 *flower_mask |= flower_flag_bit;
2302                 if (dissector_key & dissector_flag_bit)
2303                         *flower_key |= flower_flag_bit;
2304         }
2305 }
2306
2307 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2308 {
2309         u32 key, mask;
2310         __be32 _key, _mask;
2311         int err;
2312
2313         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2314                 return 0;
2315
2316         key = 0;
2317         mask = 0;
2318
2319         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2320                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2321         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2322                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2323                         FLOW_DIS_FIRST_FRAG);
2324
2325         _key = cpu_to_be32(key);
2326         _mask = cpu_to_be32(mask);
2327
2328         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2329         if (err)
2330                 return err;
2331
2332         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2333 }
2334
2335 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2336                                   struct flow_dissector_key_enc_opts *enc_opts)
2337 {
2338         struct geneve_opt *opt;
2339         struct nlattr *nest;
2340         int opt_off = 0;
2341
2342         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2343         if (!nest)
2344                 goto nla_put_failure;
2345
2346         while (enc_opts->len > opt_off) {
2347                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2348
2349                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2350                                  opt->opt_class))
2351                         goto nla_put_failure;
2352                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2353                                opt->type))
2354                         goto nla_put_failure;
2355                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2356                             opt->length * 4, opt->opt_data))
2357                         goto nla_put_failure;
2358
2359                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2360         }
2361         nla_nest_end(skb, nest);
2362         return 0;
2363
2364 nla_put_failure:
2365         nla_nest_cancel(skb, nest);
2366         return -EMSGSIZE;
2367 }
2368
2369 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2370                                  struct flow_dissector_key_enc_opts *enc_opts)
2371 {
2372         struct vxlan_metadata *md;
2373         struct nlattr *nest;
2374
2375         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2376         if (!nest)
2377                 goto nla_put_failure;
2378
2379         md = (struct vxlan_metadata *)&enc_opts->data[0];
2380         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2381                 goto nla_put_failure;
2382
2383         nla_nest_end(skb, nest);
2384         return 0;
2385
2386 nla_put_failure:
2387         nla_nest_cancel(skb, nest);
2388         return -EMSGSIZE;
2389 }
2390
2391 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2392                                   struct flow_dissector_key_enc_opts *enc_opts)
2393 {
2394         struct erspan_metadata *md;
2395         struct nlattr *nest;
2396
2397         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2398         if (!nest)
2399                 goto nla_put_failure;
2400
2401         md = (struct erspan_metadata *)&enc_opts->data[0];
2402         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2403                 goto nla_put_failure;
2404
2405         if (md->version == 1 &&
2406             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2407                 goto nla_put_failure;
2408
2409         if (md->version == 2 &&
2410             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2411                         md->u.md2.dir) ||
2412              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2413                         get_hwid(&md->u.md2))))
2414                 goto nla_put_failure;
2415
2416         nla_nest_end(skb, nest);
2417         return 0;
2418
2419 nla_put_failure:
2420         nla_nest_cancel(skb, nest);
2421         return -EMSGSIZE;
2422 }
2423
2424 static int fl_dump_key_ct(struct sk_buff *skb,
2425                           struct flow_dissector_key_ct *key,
2426                           struct flow_dissector_key_ct *mask)
2427 {
2428         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2429             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2430                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2431                             sizeof(key->ct_state)))
2432                 goto nla_put_failure;
2433
2434         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2435             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2436                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2437                             sizeof(key->ct_zone)))
2438                 goto nla_put_failure;
2439
2440         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2441             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2442                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2443                             sizeof(key->ct_mark)))
2444                 goto nla_put_failure;
2445
2446         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2447             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2448                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2449                             sizeof(key->ct_labels)))
2450                 goto nla_put_failure;
2451
2452         return 0;
2453
2454 nla_put_failure:
2455         return -EMSGSIZE;
2456 }
2457
2458 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2459                                struct flow_dissector_key_enc_opts *enc_opts)
2460 {
2461         struct nlattr *nest;
2462         int err;
2463
2464         if (!enc_opts->len)
2465                 return 0;
2466
2467         nest = nla_nest_start_noflag(skb, enc_opt_type);
2468         if (!nest)
2469                 goto nla_put_failure;
2470
2471         switch (enc_opts->dst_opt_type) {
2472         case TUNNEL_GENEVE_OPT:
2473                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2474                 if (err)
2475                         goto nla_put_failure;
2476                 break;
2477         case TUNNEL_VXLAN_OPT:
2478                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
2479                 if (err)
2480                         goto nla_put_failure;
2481                 break;
2482         case TUNNEL_ERSPAN_OPT:
2483                 err = fl_dump_key_erspan_opt(skb, enc_opts);
2484                 if (err)
2485                         goto nla_put_failure;
2486                 break;
2487         default:
2488                 goto nla_put_failure;
2489         }
2490         nla_nest_end(skb, nest);
2491         return 0;
2492
2493 nla_put_failure:
2494         nla_nest_cancel(skb, nest);
2495         return -EMSGSIZE;
2496 }
2497
2498 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2499                                struct flow_dissector_key_enc_opts *key_opts,
2500                                struct flow_dissector_key_enc_opts *msk_opts)
2501 {
2502         int err;
2503
2504         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2505         if (err)
2506                 return err;
2507
2508         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2509 }
2510
2511 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2512                        struct fl_flow_key *key, struct fl_flow_key *mask)
2513 {
2514         if (mask->meta.ingress_ifindex) {
2515                 struct net_device *dev;
2516
2517                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
2518                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2519                         goto nla_put_failure;
2520         }
2521
2522         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2523                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2524                             sizeof(key->eth.dst)) ||
2525             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2526                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2527                             sizeof(key->eth.src)) ||
2528             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2529                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2530                             sizeof(key->basic.n_proto)))
2531                 goto nla_put_failure;
2532
2533         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2534                 goto nla_put_failure;
2535
2536         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2537                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2538                 goto nla_put_failure;
2539
2540         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2541                              TCA_FLOWER_KEY_CVLAN_PRIO,
2542                              &key->cvlan, &mask->cvlan) ||
2543             (mask->cvlan.vlan_tpid &&
2544              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2545                           key->cvlan.vlan_tpid)))
2546                 goto nla_put_failure;
2547
2548         if (mask->basic.n_proto) {
2549                 if (mask->cvlan.vlan_tpid) {
2550                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
2551                                          key->basic.n_proto))
2552                                 goto nla_put_failure;
2553                 } else if (mask->vlan.vlan_tpid) {
2554                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2555                                          key->basic.n_proto))
2556                                 goto nla_put_failure;
2557                 }
2558         }
2559
2560         if ((key->basic.n_proto == htons(ETH_P_IP) ||
2561              key->basic.n_proto == htons(ETH_P_IPV6)) &&
2562             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
2563                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
2564                             sizeof(key->basic.ip_proto)) ||
2565             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
2566                 goto nla_put_failure;
2567
2568         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2569             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
2570                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
2571                              sizeof(key->ipv4.src)) ||
2572              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
2573                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
2574                              sizeof(key->ipv4.dst))))
2575                 goto nla_put_failure;
2576         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2577                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
2578                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
2579                                   sizeof(key->ipv6.src)) ||
2580                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
2581                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
2582                                   sizeof(key->ipv6.dst))))
2583                 goto nla_put_failure;
2584
2585         if (key->basic.ip_proto == IPPROTO_TCP &&
2586             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
2587                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
2588                              sizeof(key->tp.src)) ||
2589              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
2590                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
2591                              sizeof(key->tp.dst)) ||
2592              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
2593                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
2594                              sizeof(key->tcp.flags))))
2595                 goto nla_put_failure;
2596         else if (key->basic.ip_proto == IPPROTO_UDP &&
2597                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
2598                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
2599                                   sizeof(key->tp.src)) ||
2600                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
2601                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
2602                                   sizeof(key->tp.dst))))
2603                 goto nla_put_failure;
2604         else if (key->basic.ip_proto == IPPROTO_SCTP &&
2605                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
2606                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
2607                                   sizeof(key->tp.src)) ||
2608                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
2609                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
2610                                   sizeof(key->tp.dst))))
2611                 goto nla_put_failure;
2612         else if (key->basic.n_proto == htons(ETH_P_IP) &&
2613                  key->basic.ip_proto == IPPROTO_ICMP &&
2614                  (fl_dump_key_val(skb, &key->icmp.type,
2615                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
2616                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
2617                                   sizeof(key->icmp.type)) ||
2618                   fl_dump_key_val(skb, &key->icmp.code,
2619                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
2620                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
2621                                   sizeof(key->icmp.code))))
2622                 goto nla_put_failure;
2623         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
2624                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
2625                  (fl_dump_key_val(skb, &key->icmp.type,
2626                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
2627                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
2628                                   sizeof(key->icmp.type)) ||
2629                   fl_dump_key_val(skb, &key->icmp.code,
2630                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
2631                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
2632                                   sizeof(key->icmp.code))))
2633                 goto nla_put_failure;
2634         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
2635                   key->basic.n_proto == htons(ETH_P_RARP)) &&
2636                  (fl_dump_key_val(skb, &key->arp.sip,
2637                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
2638                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
2639                                   sizeof(key->arp.sip)) ||
2640                   fl_dump_key_val(skb, &key->arp.tip,
2641                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
2642                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
2643                                   sizeof(key->arp.tip)) ||
2644                   fl_dump_key_val(skb, &key->arp.op,
2645                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
2646                                   TCA_FLOWER_KEY_ARP_OP_MASK,
2647                                   sizeof(key->arp.op)) ||
2648                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2649                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2650                                   sizeof(key->arp.sha)) ||
2651                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2652                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2653                                   sizeof(key->arp.tha))))
2654                 goto nla_put_failure;
2655
2656         if ((key->basic.ip_proto == IPPROTO_TCP ||
2657              key->basic.ip_proto == IPPROTO_UDP ||
2658              key->basic.ip_proto == IPPROTO_SCTP) &&
2659              fl_dump_key_port_range(skb, key, mask))
2660                 goto nla_put_failure;
2661
2662         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2663             (fl_dump_key_val(skb, &key->enc_ipv4.src,
2664                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
2665                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
2666                             sizeof(key->enc_ipv4.src)) ||
2667              fl_dump_key_val(skb, &key->enc_ipv4.dst,
2668                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
2669                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
2670                              sizeof(key->enc_ipv4.dst))))
2671                 goto nla_put_failure;
2672         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2673                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
2674                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
2675                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
2676                             sizeof(key->enc_ipv6.src)) ||
2677                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
2678                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
2679                                  &mask->enc_ipv6.dst,
2680                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
2681                             sizeof(key->enc_ipv6.dst))))
2682                 goto nla_put_failure;
2683
2684         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
2685                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
2686                             sizeof(key->enc_key_id)) ||
2687             fl_dump_key_val(skb, &key->enc_tp.src,
2688                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
2689                             &mask->enc_tp.src,
2690                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
2691                             sizeof(key->enc_tp.src)) ||
2692             fl_dump_key_val(skb, &key->enc_tp.dst,
2693                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
2694                             &mask->enc_tp.dst,
2695                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
2696                             sizeof(key->enc_tp.dst)) ||
2697             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
2698             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
2699                 goto nla_put_failure;
2700
2701         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
2702                 goto nla_put_failure;
2703
2704         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
2705                 goto nla_put_failure;
2706
2707         return 0;
2708
2709 nla_put_failure:
2710         return -EMSGSIZE;
2711 }
2712
2713 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
2714                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
2715 {
2716         struct cls_fl_filter *f = fh;
2717         struct nlattr *nest;
2718         struct fl_flow_key *key, *mask;
2719         bool skip_hw;
2720
2721         if (!f)
2722                 return skb->len;
2723
2724         t->tcm_handle = f->handle;
2725
2726         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2727         if (!nest)
2728                 goto nla_put_failure;
2729
2730         spin_lock(&tp->lock);
2731
2732         if (f->res.classid &&
2733             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
2734                 goto nla_put_failure_locked;
2735
2736         key = &f->key;
2737         mask = &f->mask->key;
2738         skip_hw = tc_skip_hw(f->flags);
2739
2740         if (fl_dump_key(skb, net, key, mask))
2741                 goto nla_put_failure_locked;
2742
2743         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
2744                 goto nla_put_failure_locked;
2745
2746         spin_unlock(&tp->lock);
2747
2748         if (!skip_hw)
2749                 fl_hw_update_stats(tp, f, rtnl_held);
2750
2751         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
2752                 goto nla_put_failure;
2753
2754         if (tcf_exts_dump(skb, &f->exts))
2755                 goto nla_put_failure;
2756
2757         nla_nest_end(skb, nest);
2758
2759         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
2760                 goto nla_put_failure;
2761
2762         return skb->len;
2763
2764 nla_put_failure_locked:
2765         spin_unlock(&tp->lock);
2766 nla_put_failure:
2767         nla_nest_cancel(skb, nest);
2768         return -1;
2769 }
2770
2771 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
2772 {
2773         struct fl_flow_tmplt *tmplt = tmplt_priv;
2774         struct fl_flow_key *key, *mask;
2775         struct nlattr *nest;
2776
2777         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2778         if (!nest)
2779                 goto nla_put_failure;
2780
2781         key = &tmplt->dummy_key;
2782         mask = &tmplt->mask;
2783
2784         if (fl_dump_key(skb, net, key, mask))
2785                 goto nla_put_failure;
2786
2787         nla_nest_end(skb, nest);
2788
2789         return skb->len;
2790
2791 nla_put_failure:
2792         nla_nest_cancel(skb, nest);
2793         return -EMSGSIZE;
2794 }
2795
2796 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
2797                           unsigned long base)
2798 {
2799         struct cls_fl_filter *f = fh;
2800
2801         if (f && f->res.classid == classid) {
2802                 if (cl)
2803                         __tcf_bind_filter(q, &f->res, base);
2804                 else
2805                         __tcf_unbind_filter(q, &f->res);
2806         }
2807 }
2808
2809 static bool fl_delete_empty(struct tcf_proto *tp)
2810 {
2811         struct cls_fl_head *head = fl_head_dereference(tp);
2812
2813         spin_lock(&tp->lock);
2814         tp->deleting = idr_is_empty(&head->handle_idr);
2815         spin_unlock(&tp->lock);
2816
2817         return tp->deleting;
2818 }
2819
2820 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
2821         .kind           = "flower",
2822         .classify       = fl_classify,
2823         .init           = fl_init,
2824         .destroy        = fl_destroy,
2825         .get            = fl_get,
2826         .put            = fl_put,
2827         .change         = fl_change,
2828         .delete         = fl_delete,
2829         .delete_empty   = fl_delete_empty,
2830         .walk           = fl_walk,
2831         .reoffload      = fl_reoffload,
2832         .hw_add         = fl_hw_add,
2833         .hw_del         = fl_hw_del,
2834         .dump           = fl_dump,
2835         .bind_class     = fl_bind_class,
2836         .tmplt_create   = fl_tmplt_create,
2837         .tmplt_destroy  = fl_tmplt_destroy,
2838         .tmplt_dump     = fl_tmplt_dump,
2839         .owner          = THIS_MODULE,
2840         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
2841 };
2842
2843 static int __init cls_fl_init(void)
2844 {
2845         return register_tcf_proto_ops(&cls_fl_ops);
2846 }
2847
2848 static void __exit cls_fl_exit(void)
2849 {
2850         unregister_tcf_proto_ops(&cls_fl_ops);
2851 }
2852
2853 module_init(cls_fl_init);
2854 module_exit(cls_fl_exit);
2855
2856 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
2857 MODULE_DESCRIPTION("Flower classifier");
2858 MODULE_LICENSE("GPL v2");