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/pkt_sched.h>
  23 #include <net/ip.h>
  24 #include <net/flow_dissector.h>
  25 #include <net/geneve.h>
  26 #include <net/vxlan.h>
  27 #include <net/erspan.h>
  28 #include <net/gtp.h>
  29
  30 #include <net/dst.h>
  31 #include <net/dst_metadata.h>
  32
  33 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  34
  35 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  36                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  37 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  38                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  39
  40 struct fl_flow_key {
  41         struct flow_dissector_key_meta meta;
  42         struct flow_dissector_key_control control;
  43         struct flow_dissector_key_control enc_control;
  44         struct flow_dissector_key_basic basic;
  45         struct flow_dissector_key_eth_addrs eth;
  46         struct flow_dissector_key_vlan vlan;
  47         struct flow_dissector_key_vlan cvlan;
  48         union {
  49                 struct flow_dissector_key_ipv4_addrs ipv4;
  50                 struct flow_dissector_key_ipv6_addrs ipv6;
  51         };
  52         struct flow_dissector_key_ports tp;
  53         struct flow_dissector_key_icmp icmp;
  54         struct flow_dissector_key_arp arp;
  55         struct flow_dissector_key_keyid enc_key_id;
  56         union {
  57                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  58                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  59         };
  60         struct flow_dissector_key_ports enc_tp;
  61         struct flow_dissector_key_mpls mpls;
  62         struct flow_dissector_key_tcp tcp;
  63         struct flow_dissector_key_ip ip;
  64         struct flow_dissector_key_ip enc_ip;
  65         struct flow_dissector_key_enc_opts enc_opts;
  66         union {
  67                 struct flow_dissector_key_ports tp;
  68                 struct {
  69                         struct flow_dissector_key_ports tp_min;
  70                         struct flow_dissector_key_ports tp_max;
  71                 };
  72         } tp_range;
  73         struct flow_dissector_key_ct ct;
  74         struct flow_dissector_key_hash hash;
  75 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  76
  77 struct fl_flow_mask_range {
  78         unsigned short int start;
  79         unsigned short int end;
  80 };
  81
  82 struct fl_flow_mask {
  83         struct fl_flow_key key;
  84         struct fl_flow_mask_range range;
  85         u32 flags;
  86         struct rhash_head ht_node;
  87         struct rhashtable ht;
  88         struct rhashtable_params filter_ht_params;
  89         struct flow_dissector dissector;
  90         struct list_head filters;
  91         struct rcu_work rwork;
  92         struct list_head list;
  93         refcount_t refcnt;
  94 };
  95
  96 struct fl_flow_tmplt {
  97         struct fl_flow_key dummy_key;
  98         struct fl_flow_key mask;
  99         struct flow_dissector dissector;
 100         struct tcf_chain *chain;
 101 };
 102
 103 struct cls_fl_head {
 104         struct rhashtable ht;
 105         spinlock_t masks_lock; /* Protect masks list */
 106         struct list_head masks;
 107         struct list_head hw_filters;
 108         struct rcu_work rwork;
 109         struct idr handle_idr;
 110 };
 111
 112 struct cls_fl_filter {
 113         struct fl_flow_mask *mask;
 114         struct rhash_head ht_node;
 115         struct fl_flow_key mkey;
 116         struct tcf_exts exts;
 117         struct tcf_result res;
 118         struct fl_flow_key key;
 119         struct list_head list;
 120         struct list_head hw_list;
 121         u32 handle;
 122         u32 flags;
 123         u32 in_hw_count;
 124         struct rcu_work rwork;
 125         struct net_device *hw_dev;
 126         /* Flower classifier is unlocked, which means that its reference counter
 127          * can be changed concurrently without any kind of external
 128          * synchronization. Use atomic reference counter to be concurrency-safe.
 129          */
 130         refcount_t refcnt;
 131         bool deleted;
 132 };
 133
 134 static const struct rhashtable_params mask_ht_params = {
 135         .key_offset = offsetof(struct fl_flow_mask, key),
 136         .key_len = sizeof(struct fl_flow_key),
 137         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 138         .automatic_shrinking = true,
 139 };
 140
 141 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 142 {
 143         return mask->range.end - mask->range.start;
 144 }
 145
 146 static void fl_mask_update_range(struct fl_flow_mask *mask)
 147 {
 148         const u8 *bytes = (const u8 *) &mask->key;
 149         size_t size = sizeof(mask->key);
 150         size_t i, first = 0, last;
 151
 152         for (i = 0; i < size; i++) {
 153                 if (bytes[i]) {
 154                         first = i;
 155                         break;
 156                 }
 157         }
 158         last = first;
 159         for (i = size - 1; i != first; i--) {
 160                 if (bytes[i]) {
 161                         last = i;
 162                         break;
 163                 }
 164         }
 165         mask->range.start = rounddown(first, sizeof(long));
 166         mask->range.end = roundup(last + 1, sizeof(long));
 167 }
 168
 169 static void *fl_key_get_start(struct fl_flow_key *key,
 170                               const struct fl_flow_mask *mask)
 171 {
 172         return (u8 *) key + mask->range.start;
 173 }
 174
 175 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 176                               struct fl_flow_mask *mask)
 177 {
 178         const long *lkey = fl_key_get_start(key, mask);
 179         const long *lmask = fl_key_get_start(&mask->key, mask);
 180         long *lmkey = fl_key_get_start(mkey, mask);
 181         int i;
 182
 183         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 184                 *lmkey++ = *lkey++ & *lmask++;
 185 }
 186
 187 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 188                                struct fl_flow_mask *mask)
 189 {
 190         const long *lmask = fl_key_get_start(&mask->key, mask);
 191         const long *ltmplt;
 192         int i;
 193
 194         if (!tmplt)
 195                 return true;
 196         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 197         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 198                 if (~*ltmplt++ & *lmask++)
 199                         return false;
 200         }
 201         return true;
 202 }
 203
 204 static void fl_clear_masked_range(struct fl_flow_key *key,
 205                                   struct fl_flow_mask *mask)
 206 {
 207         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 208 }
 209
 210 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 211                                   struct fl_flow_key *key,
 212                                   struct fl_flow_key *mkey)
 213 {
 214         u16 min_mask, max_mask, min_val, max_val;
 215
 216         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 217         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 218         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 219         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 220
 221         if (min_mask && max_mask) {
 222                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 223                     ntohs(key->tp_range.tp.dst) > max_val)
 224                         return false;
 225
 226                 /* skb does not have min and max values */
 227                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 228                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 229         }
 230         return true;
 231 }
 232
 233 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 234                                   struct fl_flow_key *key,
 235                                   struct fl_flow_key *mkey)
 236 {
 237         u16 min_mask, max_mask, min_val, max_val;
 238
 239         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 240         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 241         min_val = ntohs(filter->key.tp_range.tp_min.src);
 242         max_val = ntohs(filter->key.tp_range.tp_max.src);
 243
 244         if (min_mask && max_mask) {
 245                 if (ntohs(key->tp_range.tp.src) < min_val ||
 246                     ntohs(key->tp_range.tp.src) > max_val)
 247                         return false;
 248
 249                 /* skb does not have min and max values */
 250                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 251                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 252         }
 253         return true;
 254 }
 255
 256 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 257                                          struct fl_flow_key *mkey)
 258 {
 259         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 260                                       mask->filter_ht_params);
 261 }
 262
 263 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 264                                              struct fl_flow_key *mkey,
 265                                              struct fl_flow_key *key)
 266 {
 267         struct cls_fl_filter *filter, *f;
 268
 269         list_for_each_entry_rcu(filter, &mask->filters, list) {
 270                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 271                         continue;
 272
 273                 if (!fl_range_port_src_cmp(filter, key, mkey))
 274                         continue;
 275
 276                 f = __fl_lookup(mask, mkey);
 277                 if (f)
 278                         return f;
 279         }
 280         return NULL;
 281 }
 282
 283 static noinline_for_stack
 284 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 285 {
 286         struct fl_flow_key mkey;
 287
 288         fl_set_masked_key(&mkey, key, mask);
 289         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 290                 return fl_lookup_range(mask, &mkey, key);
 291
 292         return __fl_lookup(mask, &mkey);
 293 }
 294
 295 static u16 fl_ct_info_to_flower_map[] = {
 296         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 297                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 298         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 299                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 300         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 301                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED |
 302                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 303         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 304                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED |
 305                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 306         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 307                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 308 };
 309
 310 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 311                        struct tcf_result *res)
 312 {
 313         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 314         bool post_ct = tc_skb_cb(skb)->post_ct;
 315         u16 zone = tc_skb_cb(skb)->zone;
 316         struct fl_flow_key skb_key;
 317         struct fl_flow_mask *mask;
 318         struct cls_fl_filter *f;
 319
 320         list_for_each_entry_rcu(mask, &head->masks, list) {
 321                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 322                 fl_clear_masked_range(&skb_key, mask);
 323
 324                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 325                 /* skb_flow_dissect() does not set n_proto in case an unknown
 326                  * protocol, so do it rather here.
 327                  */
 328                 skb_key.basic.n_proto = skb_protocol(skb, false);
 329                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 330                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 331                                     fl_ct_info_to_flower_map,
 332                                     ARRAY_SIZE(fl_ct_info_to_flower_map),
 333                                     post_ct, zone);
 334                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 335                 skb_flow_dissect(skb, &mask->dissector, &skb_key,
 336                                  FLOW_DISSECTOR_F_STOP_BEFORE_ENCAP);
 337
 338                 f = fl_mask_lookup(mask, &skb_key);
 339                 if (f && !tc_skip_sw(f->flags)) {
 340                         *res = f->res;
 341                         return tcf_exts_exec(skb, &f->exts, res);
 342                 }
 343         }
 344         return -1;
 345 }
 346
 347 static int fl_init(struct tcf_proto *tp)
 348 {
 349         struct cls_fl_head *head;
 350
 351         head = kzalloc(sizeof(*head), GFP_KERNEL);
 352         if (!head)
 353                 return -ENOBUFS;
 354
 355         spin_lock_init(&head->masks_lock);
 356         INIT_LIST_HEAD_RCU(&head->masks);
 357         INIT_LIST_HEAD(&head->hw_filters);
 358         rcu_assign_pointer(tp->root, head);
 359         idr_init(&head->handle_idr);
 360
 361         return rhashtable_init(&head->ht, &mask_ht_params);
 362 }
 363
 364 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 365 {
 366         /* temporary masks don't have their filters list and ht initialized */
 367         if (mask_init_done) {
 368                 WARN_ON(!list_empty(&mask->filters));
 369                 rhashtable_destroy(&mask->ht);
 370         }
 371         kfree(mask);
 372 }
 373
 374 static void fl_mask_free_work(struct work_struct *work)
 375 {
 376         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 377                                                  struct fl_flow_mask, rwork);
 378
 379         fl_mask_free(mask, true);
 380 }
 381
 382 static void fl_uninit_mask_free_work(struct work_struct *work)
 383 {
 384         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 385                                                  struct fl_flow_mask, rwork);
 386
 387         fl_mask_free(mask, false);
 388 }
 389
 390 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 391 {
 392         if (!refcount_dec_and_test(&mask->refcnt))
 393                 return false;
 394
 395         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 396
 397         spin_lock(&head->masks_lock);
 398         list_del_rcu(&mask->list);
 399         spin_unlock(&head->masks_lock);
 400
 401         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 402
 403         return true;
 404 }
 405
 406 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 407 {
 408         /* Flower classifier only changes root pointer during init and destroy.
 409          * Users must obtain reference to tcf_proto instance before calling its
 410          * API, so tp->root pointer is protected from concurrent call to
 411          * fl_destroy() by reference counting.
 412          */
 413         return rcu_dereference_raw(tp->root);
 414 }
 415
 416 static void __fl_destroy_filter(struct cls_fl_filter *f)
 417 {
 418         tcf_exts_destroy(&f->exts);
 419         tcf_exts_put_net(&f->exts);
 420         kfree(f);
 421 }
 422
 423 static void fl_destroy_filter_work(struct work_struct *work)
 424 {
 425         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 426                                         struct cls_fl_filter, rwork);
 427
 428         __fl_destroy_filter(f);
 429 }
 430
 431 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 432                                  bool rtnl_held, struct netlink_ext_ack *extack)
 433 {
 434         struct tcf_block *block = tp->chain->block;
 435         struct flow_cls_offload cls_flower = {};
 436
 437         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 438         cls_flower.command = FLOW_CLS_DESTROY;
 439         cls_flower.cookie = (unsigned long) f;
 440
 441         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 442                             &f->flags, &f->in_hw_count, rtnl_held);
 443
 444 }
 445
 446 static int fl_hw_replace_filter(struct tcf_proto *tp,
 447                                 struct cls_fl_filter *f, bool rtnl_held,
 448                                 struct netlink_ext_ack *extack)
 449 {
 450         struct tcf_block *block = tp->chain->block;
 451         struct flow_cls_offload cls_flower = {};
 452         bool skip_sw = tc_skip_sw(f->flags);
 453         int err = 0;
 454
 455         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 456         if (!cls_flower.rule)
 457                 return -ENOMEM;
 458
 459         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 460         cls_flower.command = FLOW_CLS_REPLACE;
 461         cls_flower.cookie = (unsigned long) f;
 462         cls_flower.rule->match.dissector = &f->mask->dissector;
 463         cls_flower.rule->match.mask = &f->mask->key;
 464         cls_flower.rule->match.key = &f->mkey;
 465         cls_flower.classid = f->res.classid;
 466
 467         err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts);
 468         if (err) {
 469                 kfree(cls_flower.rule);
 470                 if (skip_sw) {
 471                         NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
 472                         return err;
 473                 }
 474                 return 0;
 475         }
 476
 477         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 478                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 479         tc_cleanup_offload_action(&cls_flower.rule->action);
 480         kfree(cls_flower.rule);
 481
 482         if (err) {
 483                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 484                 return err;
 485         }
 486
 487         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 488                 return -EINVAL;
 489
 490         return 0;
 491 }
 492
 493 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 494                                bool rtnl_held)
 495 {
 496         struct tcf_block *block = tp->chain->block;
 497         struct flow_cls_offload cls_flower = {};
 498
 499         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 500         cls_flower.command = FLOW_CLS_STATS;
 501         cls_flower.cookie = (unsigned long) f;
 502         cls_flower.classid = f->res.classid;
 503
 504         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 505                          rtnl_held);
 506
 507         tcf_exts_hw_stats_update(&f->exts, cls_flower.stats.bytes,
 508                                  cls_flower.stats.pkts,
 509                                  cls_flower.stats.drops,
 510                                  cls_flower.stats.lastused,
 511                                  cls_flower.stats.used_hw_stats,
 512                                  cls_flower.stats.used_hw_stats_valid);
 513 }
 514
 515 static void __fl_put(struct cls_fl_filter *f)
 516 {
 517         if (!refcount_dec_and_test(&f->refcnt))
 518                 return;
 519
 520         if (tcf_exts_get_net(&f->exts))
 521                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 522         else
 523                 __fl_destroy_filter(f);
 524 }
 525
 526 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 527 {
 528         struct cls_fl_filter *f;
 529
 530         rcu_read_lock();
 531         f = idr_find(&head->handle_idr, handle);
 532         if (f && !refcount_inc_not_zero(&f->refcnt))
 533                 f = NULL;
 534         rcu_read_unlock();
 535
 536         return f;
 537 }
 538
 539 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 540                        bool *last, bool rtnl_held,
 541                        struct netlink_ext_ack *extack)
 542 {
 543         struct cls_fl_head *head = fl_head_dereference(tp);
 544
 545         *last = false;
 546
 547         spin_lock(&tp->lock);
 548         if (f->deleted) {
 549                 spin_unlock(&tp->lock);
 550                 return -ENOENT;
 551         }
 552
 553         f->deleted = true;
 554         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 555                                f->mask->filter_ht_params);
 556         idr_remove(&head->handle_idr, f->handle);
 557         list_del_rcu(&f->list);
 558         spin_unlock(&tp->lock);
 559
 560         *last = fl_mask_put(head, f->mask);
 561         if (!tc_skip_hw(f->flags))
 562                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 563         tcf_unbind_filter(tp, &f->res);
 564         __fl_put(f);
 565
 566         return 0;
 567 }
 568
 569 static void fl_destroy_sleepable(struct work_struct *work)
 570 {
 571         struct cls_fl_head *head = container_of(to_rcu_work(work),
 572                                                 struct cls_fl_head,
 573                                                 rwork);
 574
 575         rhashtable_destroy(&head->ht);
 576         kfree(head);
 577         module_put(THIS_MODULE);
 578 }
 579
 580 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 581                        struct netlink_ext_ack *extack)
 582 {
 583         struct cls_fl_head *head = fl_head_dereference(tp);
 584         struct fl_flow_mask *mask, *next_mask;
 585         struct cls_fl_filter *f, *next;
 586         bool last;
 587
 588         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 589                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 590                         __fl_delete(tp, f, &last, rtnl_held, extack);
 591                         if (last)
 592                                 break;
 593                 }
 594         }
 595         idr_destroy(&head->handle_idr);
 596
 597         __module_get(THIS_MODULE);
 598         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 599 }
 600
 601 static void fl_put(struct tcf_proto *tp, void *arg)
 602 {
 603         struct cls_fl_filter *f = arg;
 604
 605         __fl_put(f);
 606 }
 607
 608 static void *fl_get(struct tcf_proto *tp, u32 handle)
 609 {
 610         struct cls_fl_head *head = fl_head_dereference(tp);
 611
 612         return __fl_get(head, handle);
 613 }
 614
 615 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 616         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 617         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 618         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 619                                             .len = IFNAMSIZ },
 620         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 621         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 622         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 623         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 624         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 625         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 626         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 627         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 628         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 629         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 630         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 631         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 632         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 633         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 634         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 635         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 636         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 637         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 638         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 639         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 640         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 641         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 642         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 643         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 644         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 645         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 646         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 647         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 648         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 649         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 650         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 660         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 661         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 662         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 663         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 664         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 665         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 666         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 670         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 671         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 672         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 673         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 674         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 675         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 676         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 677         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 678         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 679         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 680         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 681         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 682         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 683         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 684         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 685         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 686         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 687         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 688         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 689         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 690         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 692         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 693         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 694         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 695         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 696         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 697         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 698         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 699         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 700         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 701         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 702         [TCA_FLOWER_KEY_CT_STATE]       =
 703                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 704         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 705                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 706         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 707         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 708         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 709         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 710         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 711                                             .len = 128 / BITS_PER_BYTE },
 712         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 713                                             .len = 128 / BITS_PER_BYTE },
 714         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 715         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 716         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 717
 718 };
 719
 720 static const struct nla_policy
 721 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 722         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 723                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 724         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 725         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 726         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 727         [TCA_FLOWER_KEY_ENC_OPTS_GTP]           = { .type = NLA_NESTED },
 728 };
 729
 730 static const struct nla_policy
 731 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 732         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 733         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 734         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 735                                                        .len = 128 },
 736 };
 737
 738 static const struct nla_policy
 739 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 740         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 741 };
 742
 743 static const struct nla_policy
 744 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 745         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 746         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 747         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 748         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 749 };
 750
 751 static const struct nla_policy
 752 gtp_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1] = {
 753         [TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]      = { .type = NLA_U8 },
 754         [TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]           = { .type = NLA_U8 },
 755 };
 756
 757 static const struct nla_policy
 758 mpls_stack_entry_policy[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1] = {
 759         [TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]    = { .type = NLA_U8 },
 760         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]      = { .type = NLA_U8 },
 761         [TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]      = { .type = NLA_U8 },
 762         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]       = { .type = NLA_U8 },
 763         [TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]    = { .type = NLA_U32 },
 764 };
 765
 766 static void fl_set_key_val(struct nlattr **tb,
 767                            void *val, int val_type,
 768                            void *mask, int mask_type, int len)
 769 {
 770         if (!tb[val_type])
 771                 return;
 772         nla_memcpy(val, tb[val_type], len);
 773         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 774                 memset(mask, 0xff, len);
 775         else
 776                 nla_memcpy(mask, tb[mask_type], len);
 777 }
 778
 779 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 780                                  struct fl_flow_key *mask,
 781                                  struct netlink_ext_ack *extack)
 782 {
 783         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 784                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 785                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 786         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 787                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 788                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 789         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 790                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 791                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 792         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 793                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 794                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 795
 796         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 797             ntohs(key->tp_range.tp_max.dst) <=
 798             ntohs(key->tp_range.tp_min.dst)) {
 799                 NL_SET_ERR_MSG_ATTR(extack,
 800                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 801                                     "Invalid destination port range (min must be strictly smaller than max)");
 802                 return -EINVAL;
 803         }
 804         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 805             ntohs(key->tp_range.tp_max.src) <=
 806             ntohs(key->tp_range.tp_min.src)) {
 807                 NL_SET_ERR_MSG_ATTR(extack,
 808                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 809                                     "Invalid source port range (min must be strictly smaller than max)");
 810                 return -EINVAL;
 811         }
 812
 813         return 0;
 814 }
 815
 816 static int fl_set_key_mpls_lse(const struct nlattr *nla_lse,
 817                                struct flow_dissector_key_mpls *key_val,
 818                                struct flow_dissector_key_mpls *key_mask,
 819                                struct netlink_ext_ack *extack)
 820 {
 821         struct nlattr *tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1];
 822         struct flow_dissector_mpls_lse *lse_mask;
 823         struct flow_dissector_mpls_lse *lse_val;
 824         u8 lse_index;
 825         u8 depth;
 826         int err;
 827
 828         err = nla_parse_nested(tb, TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX, nla_lse,
 829                                mpls_stack_entry_policy, extack);
 830         if (err < 0)
 831                 return err;
 832
 833         if (!tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]) {
 834                 NL_SET_ERR_MSG(extack, "Missing MPLS option \"depth\"");
 835                 return -EINVAL;
 836         }
 837
 838         depth = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]);
 839
 840         /* LSE depth starts at 1, for consistency with terminology used by
 841          * RFC 3031 (section 3.9), where depth 0 refers to unlabeled packets.
 842          */
 843         if (depth < 1 || depth > FLOW_DIS_MPLS_MAX) {
 844                 NL_SET_ERR_MSG_ATTR(extack,
 845                                     tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH],
 846                                     "Invalid MPLS depth");
 847                 return -EINVAL;
 848         }
 849         lse_index = depth - 1;
 850
 851         dissector_set_mpls_lse(key_val, lse_index);
 852         dissector_set_mpls_lse(key_mask, lse_index);
 853
 854         lse_val = &key_val->ls[lse_index];
 855         lse_mask = &key_mask->ls[lse_index];
 856
 857         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]) {
 858                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]);
 859                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 860         }
 861         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]) {
 862                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]);
 863
 864                 if (bos & ~MPLS_BOS_MASK) {
 865                         NL_SET_ERR_MSG_ATTR(extack,
 866                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS],
 867                                             "Bottom Of Stack (BOS) must be 0 or 1");
 868                         return -EINVAL;
 869                 }
 870                 lse_val->mpls_bos = bos;
 871                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 872         }
 873         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]) {
 874                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]);
 875
 876                 if (tc & ~MPLS_TC_MASK) {
 877                         NL_SET_ERR_MSG_ATTR(extack,
 878                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC],
 879                                             "Traffic Class (TC) must be between 0 and 7");
 880                         return -EINVAL;
 881                 }
 882                 lse_val->mpls_tc = tc;
 883                 lse_mask->mpls_tc = MPLS_TC_MASK;
 884         }
 885         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]) {
 886                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]);
 887
 888                 if (label & ~MPLS_LABEL_MASK) {
 889                         NL_SET_ERR_MSG_ATTR(extack,
 890                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL],
 891                                             "Label must be between 0 and 1048575");
 892                         return -EINVAL;
 893                 }
 894                 lse_val->mpls_label = label;
 895                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 896         }
 897
 898         return 0;
 899 }
 900
 901 static int fl_set_key_mpls_opts(const struct nlattr *nla_mpls_opts,
 902                                 struct flow_dissector_key_mpls *key_val,
 903                                 struct flow_dissector_key_mpls *key_mask,
 904                                 struct netlink_ext_ack *extack)
 905 {
 906         struct nlattr *nla_lse;
 907         int rem;
 908         int err;
 909
 910         if (!(nla_mpls_opts->nla_type & NLA_F_NESTED)) {
 911                 NL_SET_ERR_MSG_ATTR(extack, nla_mpls_opts,
 912                                     "NLA_F_NESTED is missing");
 913                 return -EINVAL;
 914         }
 915
 916         nla_for_each_nested(nla_lse, nla_mpls_opts, rem) {
 917                 if (nla_type(nla_lse) != TCA_FLOWER_KEY_MPLS_OPTS_LSE) {
 918                         NL_SET_ERR_MSG_ATTR(extack, nla_lse,
 919                                             "Invalid MPLS option type");
 920                         return -EINVAL;
 921                 }
 922
 923                 err = fl_set_key_mpls_lse(nla_lse, key_val, key_mask, extack);
 924                 if (err < 0)
 925                         return err;
 926         }
 927         if (rem) {
 928                 NL_SET_ERR_MSG(extack,
 929                                "Bytes leftover after parsing MPLS options");
 930                 return -EINVAL;
 931         }
 932
 933         return 0;
 934 }
 935
 936 static int fl_set_key_mpls(struct nlattr **tb,
 937                            struct flow_dissector_key_mpls *key_val,
 938                            struct flow_dissector_key_mpls *key_mask,
 939                            struct netlink_ext_ack *extack)
 940 {
 941         struct flow_dissector_mpls_lse *lse_mask;
 942         struct flow_dissector_mpls_lse *lse_val;
 943
 944         if (tb[TCA_FLOWER_KEY_MPLS_OPTS]) {
 945                 if (tb[TCA_FLOWER_KEY_MPLS_TTL] ||
 946                     tb[TCA_FLOWER_KEY_MPLS_BOS] ||
 947                     tb[TCA_FLOWER_KEY_MPLS_TC] ||
 948                     tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 949                         NL_SET_ERR_MSG_ATTR(extack,
 950                                             tb[TCA_FLOWER_KEY_MPLS_OPTS],
 951                                             "MPLS label, Traffic Class, Bottom Of Stack and Time To Live must be encapsulated in the MPLS options attribute");
 952                         return -EBADMSG;
 953                 }
 954
 955                 return fl_set_key_mpls_opts(tb[TCA_FLOWER_KEY_MPLS_OPTS],
 956                                             key_val, key_mask, extack);
 957         }
 958
 959         lse_val = &key_val->ls[0];
 960         lse_mask = &key_mask->ls[0];
 961
 962         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 963                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 964                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 965                 dissector_set_mpls_lse(key_val, 0);
 966                 dissector_set_mpls_lse(key_mask, 0);
 967         }
 968         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 969                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 970
 971                 if (bos & ~MPLS_BOS_MASK) {
 972                         NL_SET_ERR_MSG_ATTR(extack,
 973                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
 974                                             "Bottom Of Stack (BOS) must be 0 or 1");
 975                         return -EINVAL;
 976                 }
 977                 lse_val->mpls_bos = bos;
 978                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 979                 dissector_set_mpls_lse(key_val, 0);
 980                 dissector_set_mpls_lse(key_mask, 0);
 981         }
 982         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 983                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 984
 985                 if (tc & ~MPLS_TC_MASK) {
 986                         NL_SET_ERR_MSG_ATTR(extack,
 987                                             tb[TCA_FLOWER_KEY_MPLS_TC],
 988                                             "Traffic Class (TC) must be between 0 and 7");
 989                         return -EINVAL;
 990                 }
 991                 lse_val->mpls_tc = tc;
 992                 lse_mask->mpls_tc = MPLS_TC_MASK;
 993                 dissector_set_mpls_lse(key_val, 0);
 994                 dissector_set_mpls_lse(key_mask, 0);
 995         }
 996         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 997                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 998
 999                 if (label & ~MPLS_LABEL_MASK) {
1000                         NL_SET_ERR_MSG_ATTR(extack,
1001                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
1002                                             "Label must be between 0 and 1048575");
1003                         return -EINVAL;
1004                 }
1005                 lse_val->mpls_label = label;
1006                 lse_mask->mpls_label = MPLS_LABEL_MASK;
1007                 dissector_set_mpls_lse(key_val, 0);
1008                 dissector_set_mpls_lse(key_mask, 0);
1009         }
1010         return 0;
1011 }
1012
1013 static void fl_set_key_vlan(struct nlattr **tb,
1014                             __be16 ethertype,
1015                             int vlan_id_key, int vlan_prio_key,
1016                             struct flow_dissector_key_vlan *key_val,
1017                             struct flow_dissector_key_vlan *key_mask)
1018 {
1019 #define VLAN_PRIORITY_MASK      0x7
1020
1021         if (tb[vlan_id_key]) {
1022                 key_val->vlan_id =
1023                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1024                 key_mask->vlan_id = VLAN_VID_MASK;
1025         }
1026         if (tb[vlan_prio_key]) {
1027                 key_val->vlan_priority =
1028                         nla_get_u8(tb[vlan_prio_key]) &
1029                         VLAN_PRIORITY_MASK;
1030                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1031         }
1032         key_val->vlan_tpid = ethertype;
1033         key_mask->vlan_tpid = cpu_to_be16(~0);
1034 }
1035
1036 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1037                             u32 *dissector_key, u32 *dissector_mask,
1038                             u32 flower_flag_bit, u32 dissector_flag_bit)
1039 {
1040         if (flower_mask & flower_flag_bit) {
1041                 *dissector_mask |= dissector_flag_bit;
1042                 if (flower_key & flower_flag_bit)
1043                         *dissector_key |= dissector_flag_bit;
1044         }
1045 }
1046
1047 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1048                             u32 *flags_mask, struct netlink_ext_ack *extack)
1049 {
1050         u32 key, mask;
1051
1052         /* mask is mandatory for flags */
1053         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1054                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1055                 return -EINVAL;
1056         }
1057
1058         key = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS]));
1059         mask = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1060
1061         *flags_key  = 0;
1062         *flags_mask = 0;
1063
1064         fl_set_key_flag(key, mask, flags_key, flags_mask,
1065                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1066         fl_set_key_flag(key, mask, flags_key, flags_mask,
1067                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1068                         FLOW_DIS_FIRST_FRAG);
1069
1070         return 0;
1071 }
1072
1073 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1074                           struct flow_dissector_key_ip *key,
1075                           struct flow_dissector_key_ip *mask)
1076 {
1077         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1078         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1079         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1080         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1081
1082         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1083         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1084 }
1085
1086 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1087                              int depth, int option_len,
1088                              struct netlink_ext_ack *extack)
1089 {
1090         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1091         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1092         struct geneve_opt *opt;
1093         int err, data_len = 0;
1094
1095         if (option_len > sizeof(struct geneve_opt))
1096                 data_len = option_len - sizeof(struct geneve_opt);
1097
1098         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1099         memset(opt, 0xff, option_len);
1100         opt->length = data_len / 4;
1101         opt->r1 = 0;
1102         opt->r2 = 0;
1103         opt->r3 = 0;
1104
1105         /* If no mask has been prodived we assume an exact match. */
1106         if (!depth)
1107                 return sizeof(struct geneve_opt) + data_len;
1108
1109         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1110                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1111                 return -EINVAL;
1112         }
1113
1114         err = nla_parse_nested_deprecated(tb,
1115                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1116                                           nla, geneve_opt_policy, extack);
1117         if (err < 0)
1118                 return err;
1119
1120         /* We are not allowed to omit any of CLASS, TYPE or DATA
1121          * fields from the key.
1122          */
1123         if (!option_len &&
1124             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1125              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1126              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1127                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1128                 return -EINVAL;
1129         }
1130
1131         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1132          * for the mask.
1133          */
1134         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1135                 int new_len = key->enc_opts.len;
1136
1137                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1138                 data_len = nla_len(data);
1139                 if (data_len < 4) {
1140                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1141                         return -ERANGE;
1142                 }
1143                 if (data_len % 4) {
1144                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1145                         return -ERANGE;
1146                 }
1147
1148                 new_len += sizeof(struct geneve_opt) + data_len;
1149                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1150                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1151                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1152                         return -ERANGE;
1153                 }
1154                 opt->length = data_len / 4;
1155                 memcpy(opt->opt_data, nla_data(data), data_len);
1156         }
1157
1158         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1159                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1160                 opt->opt_class = nla_get_be16(class);
1161         }
1162
1163         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1164                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1165                 opt->type = nla_get_u8(type);
1166         }
1167
1168         return sizeof(struct geneve_opt) + data_len;
1169 }
1170
1171 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1172                             int depth, int option_len,
1173                             struct netlink_ext_ack *extack)
1174 {
1175         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1176         struct vxlan_metadata *md;
1177         int err;
1178
1179         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1180         memset(md, 0xff, sizeof(*md));
1181
1182         if (!depth)
1183                 return sizeof(*md);
1184
1185         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1186                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1187                 return -EINVAL;
1188         }
1189
1190         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1191                                vxlan_opt_policy, extack);
1192         if (err < 0)
1193                 return err;
1194
1195         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1196                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1197                 return -EINVAL;
1198         }
1199
1200         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1201                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1202                 md->gbp &= VXLAN_GBP_MASK;
1203         }
1204
1205         return sizeof(*md);
1206 }
1207
1208 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1209                              int depth, int option_len,
1210                              struct netlink_ext_ack *extack)
1211 {
1212         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1213         struct erspan_metadata *md;
1214         int err;
1215
1216         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1217         memset(md, 0xff, sizeof(*md));
1218         md->version = 1;
1219
1220         if (!depth)
1221                 return sizeof(*md);
1222
1223         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1224                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1225                 return -EINVAL;
1226         }
1227
1228         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1229                                erspan_opt_policy, extack);
1230         if (err < 0)
1231                 return err;
1232
1233         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1234                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1235                 return -EINVAL;
1236         }
1237
1238         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1239                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1240
1241         if (md->version == 1) {
1242                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1243                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1244                         return -EINVAL;
1245                 }
1246                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1247                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1248                         memset(&md->u, 0x00, sizeof(md->u));
1249                         md->u.index = nla_get_be32(nla);
1250                 }
1251         } else if (md->version == 2) {
1252                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1253                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1254                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1255                         return -EINVAL;
1256                 }
1257                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1258                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1259                         md->u.md2.dir = nla_get_u8(nla);
1260                 }
1261                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1262                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1263                         set_hwid(&md->u.md2, nla_get_u8(nla));
1264                 }
1265         } else {
1266                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1267                 return -EINVAL;
1268         }
1269
1270         return sizeof(*md);
1271 }
1272
1273 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1274                           int depth, int option_len,
1275                           struct netlink_ext_ack *extack)
1276 {
1277         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1278         struct gtp_pdu_session_info *sinfo;
1279         u8 len = key->enc_opts.len;
1280         int err;
1281
1282         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1283         memset(sinfo, 0xff, option_len);
1284
1285         if (!depth)
1286                 return sizeof(*sinfo);
1287
1288         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1289                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1290                 return -EINVAL;
1291         }
1292
1293         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1294                                gtp_opt_policy, extack);
1295         if (err < 0)
1296                 return err;
1297
1298         if (!option_len &&
1299             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1300              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1301                 NL_SET_ERR_MSG_MOD(extack,
1302                                    "Missing tunnel key gtp option pdu type or qfi");
1303                 return -EINVAL;
1304         }
1305
1306         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1307                 sinfo->pdu_type =
1308                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1309
1310         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1311                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1312
1313         return sizeof(*sinfo);
1314 }
1315
1316 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1317                           struct fl_flow_key *mask,
1318                           struct netlink_ext_ack *extack)
1319 {
1320         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1321         int err, option_len, key_depth, msk_depth = 0;
1322
1323         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1324                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1325                                              enc_opts_policy, extack);
1326         if (err)
1327                 return err;
1328
1329         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1330
1331         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1332                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1333                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1334                                                      enc_opts_policy, extack);
1335                 if (err)
1336                         return err;
1337
1338                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1339                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1340                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1341                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1342                         return -EINVAL;
1343                 }
1344         }
1345
1346         nla_for_each_attr(nla_opt_key, nla_enc_key,
1347                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1348                 switch (nla_type(nla_opt_key)) {
1349                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1350                         if (key->enc_opts.dst_opt_type &&
1351                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1352                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1353                                 return -EINVAL;
1354                         }
1355                         option_len = 0;
1356                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1357                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1358                                                        key_depth, option_len,
1359                                                        extack);
1360                         if (option_len < 0)
1361                                 return option_len;
1362
1363                         key->enc_opts.len += option_len;
1364                         /* At the same time we need to parse through the mask
1365                          * in order to verify exact and mask attribute lengths.
1366                          */
1367                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1368                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1369                                                        msk_depth, option_len,
1370                                                        extack);
1371                         if (option_len < 0)
1372                                 return option_len;
1373
1374                         mask->enc_opts.len += option_len;
1375                         if (key->enc_opts.len != mask->enc_opts.len) {
1376                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1377                                 return -EINVAL;
1378                         }
1379                         break;
1380                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1381                         if (key->enc_opts.dst_opt_type) {
1382                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1383                                 return -EINVAL;
1384                         }
1385                         option_len = 0;
1386                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1387                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1388                                                       key_depth, option_len,
1389                                                       extack);
1390                         if (option_len < 0)
1391                                 return option_len;
1392
1393                         key->enc_opts.len += option_len;
1394                         /* At the same time we need to parse through the mask
1395                          * in order to verify exact and mask attribute lengths.
1396                          */
1397                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1398                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1399                                                       msk_depth, option_len,
1400                                                       extack);
1401                         if (option_len < 0)
1402                                 return option_len;
1403
1404                         mask->enc_opts.len += option_len;
1405                         if (key->enc_opts.len != mask->enc_opts.len) {
1406                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1407                                 return -EINVAL;
1408                         }
1409                         break;
1410                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1411                         if (key->enc_opts.dst_opt_type) {
1412                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1413                                 return -EINVAL;
1414                         }
1415                         option_len = 0;
1416                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1417                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1418                                                        key_depth, option_len,
1419                                                        extack);
1420                         if (option_len < 0)
1421                                 return option_len;
1422
1423                         key->enc_opts.len += option_len;
1424                         /* At the same time we need to parse through the mask
1425                          * in order to verify exact and mask attribute lengths.
1426                          */
1427                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1428                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1429                                                        msk_depth, option_len,
1430                                                        extack);
1431                         if (option_len < 0)
1432                                 return option_len;
1433
1434                         mask->enc_opts.len += option_len;
1435                         if (key->enc_opts.len != mask->enc_opts.len) {
1436                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1437                                 return -EINVAL;
1438                         }
1439                         break;
1440                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1441                         if (key->enc_opts.dst_opt_type) {
1442                                 NL_SET_ERR_MSG_MOD(extack,
1443                                                    "Duplicate type for gtp options");
1444                                 return -EINVAL;
1445                         }
1446                         option_len = 0;
1447                         key->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1448                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1449                                                     key_depth, option_len,
1450                                                     extack);
1451                         if (option_len < 0)
1452                                 return option_len;
1453
1454                         key->enc_opts.len += option_len;
1455                         /* At the same time we need to parse through the mask
1456                          * in order to verify exact and mask attribute lengths.
1457                          */
1458                         mask->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1459                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1460                                                     msk_depth, option_len,
1461                                                     extack);
1462                         if (option_len < 0)
1463                                 return option_len;
1464
1465                         mask->enc_opts.len += option_len;
1466                         if (key->enc_opts.len != mask->enc_opts.len) {
1467                                 NL_SET_ERR_MSG_MOD(extack,
1468                                                    "Key and mask miss aligned");
1469                                 return -EINVAL;
1470                         }
1471                         break;
1472                 default:
1473                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1474                         return -EINVAL;
1475                 }
1476
1477                 if (!msk_depth)
1478                         continue;
1479
1480                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1481                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1482                         return -EINVAL;
1483                 }
1484                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1485         }
1486
1487         return 0;
1488 }
1489
1490 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1491                                 struct netlink_ext_ack *extack)
1492 {
1493         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1494                 NL_SET_ERR_MSG_ATTR(extack, tb,
1495                                     "no trk, so no other flag can be set");
1496                 return -EINVAL;
1497         }
1498
1499         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1500             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1501                 NL_SET_ERR_MSG_ATTR(extack, tb,
1502                                     "new and est are mutually exclusive");
1503                 return -EINVAL;
1504         }
1505
1506         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1507             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1508                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1509                 NL_SET_ERR_MSG_ATTR(extack, tb,
1510                                     "when inv is set, only trk may be set");
1511                 return -EINVAL;
1512         }
1513
1514         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1515             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1516                 NL_SET_ERR_MSG_ATTR(extack, tb,
1517                                     "new and rpl are mutually exclusive");
1518                 return -EINVAL;
1519         }
1520
1521         return 0;
1522 }
1523
1524 static int fl_set_key_ct(struct nlattr **tb,
1525                          struct flow_dissector_key_ct *key,
1526                          struct flow_dissector_key_ct *mask,
1527                          struct netlink_ext_ack *extack)
1528 {
1529         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1530                 int err;
1531
1532                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1533                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1534                         return -EOPNOTSUPP;
1535                 }
1536                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1537                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1538                                sizeof(key->ct_state));
1539
1540                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1541                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1542                                            extack);
1543                 if (err)
1544                         return err;
1545
1546         }
1547         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1548                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1549                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1550                         return -EOPNOTSUPP;
1551                 }
1552                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1553                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1554                                sizeof(key->ct_zone));
1555         }
1556         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1557                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1558                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1559                         return -EOPNOTSUPP;
1560                 }
1561                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1562                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1563                                sizeof(key->ct_mark));
1564         }
1565         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1566                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1567                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1568                         return -EOPNOTSUPP;
1569                 }
1570                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1571                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1572                                sizeof(key->ct_labels));
1573         }
1574
1575         return 0;
1576 }
1577
1578 static int fl_set_key(struct net *net, struct nlattr **tb,
1579                       struct fl_flow_key *key, struct fl_flow_key *mask,
1580                       struct netlink_ext_ack *extack)
1581 {
1582         __be16 ethertype;
1583         int ret = 0;
1584
1585         if (tb[TCA_FLOWER_INDEV]) {
1586                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1587                 if (err < 0)
1588                         return err;
1589                 key->meta.ingress_ifindex = err;
1590                 mask->meta.ingress_ifindex = 0xffffffff;
1591         }
1592
1593         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1594                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1595                        sizeof(key->eth.dst));
1596         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1597                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1598                        sizeof(key->eth.src));
1599
1600         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1601                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1602
1603                 if (eth_type_vlan(ethertype)) {
1604                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1605                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
1606                                         &mask->vlan);
1607
1608                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1609                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1610                                 if (eth_type_vlan(ethertype)) {
1611                                         fl_set_key_vlan(tb, ethertype,
1612                                                         TCA_FLOWER_KEY_CVLAN_ID,
1613                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1614                                                         &key->cvlan, &mask->cvlan);
1615                                         fl_set_key_val(tb, &key->basic.n_proto,
1616                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1617                                                        &mask->basic.n_proto,
1618                                                        TCA_FLOWER_UNSPEC,
1619                                                        sizeof(key->basic.n_proto));
1620                                 } else {
1621                                         key->basic.n_proto = ethertype;
1622                                         mask->basic.n_proto = cpu_to_be16(~0);
1623                                 }
1624                         }
1625                 } else {
1626                         key->basic.n_proto = ethertype;
1627                         mask->basic.n_proto = cpu_to_be16(~0);
1628                 }
1629         }
1630
1631         if (key->basic.n_proto == htons(ETH_P_IP) ||
1632             key->basic.n_proto == htons(ETH_P_IPV6)) {
1633                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1634                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1635                                sizeof(key->basic.ip_proto));
1636                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1637         }
1638
1639         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1640                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1641                 mask->control.addr_type = ~0;
1642                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1643                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1644                                sizeof(key->ipv4.src));
1645                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1646                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1647                                sizeof(key->ipv4.dst));
1648         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1649                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1650                 mask->control.addr_type = ~0;
1651                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1652                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1653                                sizeof(key->ipv6.src));
1654                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1655                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1656                                sizeof(key->ipv6.dst));
1657         }
1658
1659         if (key->basic.ip_proto == IPPROTO_TCP) {
1660                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1661                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1662                                sizeof(key->tp.src));
1663                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1664                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1665                                sizeof(key->tp.dst));
1666                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1667                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1668                                sizeof(key->tcp.flags));
1669         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1670                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1671                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1672                                sizeof(key->tp.src));
1673                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1674                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1675                                sizeof(key->tp.dst));
1676         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1677                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1678                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1679                                sizeof(key->tp.src));
1680                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1681                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1682                                sizeof(key->tp.dst));
1683         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1684                    key->basic.ip_proto == IPPROTO_ICMP) {
1685                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1686                                &mask->icmp.type,
1687                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1688                                sizeof(key->icmp.type));
1689                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1690                                &mask->icmp.code,
1691                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1692                                sizeof(key->icmp.code));
1693         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1694                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1695                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1696                                &mask->icmp.type,
1697                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1698                                sizeof(key->icmp.type));
1699                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1700                                &mask->icmp.code,
1701                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1702                                sizeof(key->icmp.code));
1703         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1704                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1705                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1706                 if (ret)
1707                         return ret;
1708         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1709                    key->basic.n_proto == htons(ETH_P_RARP)) {
1710                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1711                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1712                                sizeof(key->arp.sip));
1713                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1714                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1715                                sizeof(key->arp.tip));
1716                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1717                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1718                                sizeof(key->arp.op));
1719                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1720                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1721                                sizeof(key->arp.sha));
1722                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1723                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1724                                sizeof(key->arp.tha));
1725         }
1726
1727         if (key->basic.ip_proto == IPPROTO_TCP ||
1728             key->basic.ip_proto == IPPROTO_UDP ||
1729             key->basic.ip_proto == IPPROTO_SCTP) {
1730                 ret = fl_set_key_port_range(tb, key, mask, extack);
1731                 if (ret)
1732                         return ret;
1733         }
1734
1735         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1736             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1737                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1738                 mask->enc_control.addr_type = ~0;
1739                 fl_set_key_val(tb, &key->enc_ipv4.src,
1740                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1741                                &mask->enc_ipv4.src,
1742                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1743                                sizeof(key->enc_ipv4.src));
1744                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1745                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1746                                &mask->enc_ipv4.dst,
1747                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1748                                sizeof(key->enc_ipv4.dst));
1749         }
1750
1751         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1752             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1753                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1754                 mask->enc_control.addr_type = ~0;
1755                 fl_set_key_val(tb, &key->enc_ipv6.src,
1756                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1757                                &mask->enc_ipv6.src,
1758                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1759                                sizeof(key->enc_ipv6.src));
1760                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1761                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1762                                &mask->enc_ipv6.dst,
1763                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1764                                sizeof(key->enc_ipv6.dst));
1765         }
1766
1767         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1768                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1769                        sizeof(key->enc_key_id.keyid));
1770
1771         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1772                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1773                        sizeof(key->enc_tp.src));
1774
1775         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1776                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1777                        sizeof(key->enc_tp.dst));
1778
1779         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1780
1781         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1782                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1783                        sizeof(key->hash.hash));
1784
1785         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1786                 ret = fl_set_enc_opt(tb, key, mask, extack);
1787                 if (ret)
1788                         return ret;
1789         }
1790
1791         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1792         if (ret)
1793                 return ret;
1794
1795         if (tb[TCA_FLOWER_KEY_FLAGS])
1796                 ret = fl_set_key_flags(tb, &key->control.flags,
1797                                        &mask->control.flags, extack);
1798
1799         return ret;
1800 }
1801
1802 static void fl_mask_copy(struct fl_flow_mask *dst,
1803                          struct fl_flow_mask *src)
1804 {
1805         const void *psrc = fl_key_get_start(&src->key, src);
1806         void *pdst = fl_key_get_start(&dst->key, src);
1807
1808         memcpy(pdst, psrc, fl_mask_range(src));
1809         dst->range = src->range;
1810 }
1811
1812 static const struct rhashtable_params fl_ht_params = {
1813         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1814         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1815         .automatic_shrinking = true,
1816 };
1817
1818 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1819 {
1820         mask->filter_ht_params = fl_ht_params;
1821         mask->filter_ht_params.key_len = fl_mask_range(mask);
1822         mask->filter_ht_params.key_offset += mask->range.start;
1823
1824         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1825 }
1826
1827 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1828 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1829
1830 #define FL_KEY_IS_MASKED(mask, member)                                          \
1831         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1832                    0, FL_KEY_MEMBER_SIZE(member))                               \
1833
1834 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1835         do {                                                                    \
1836                 keys[cnt].key_id = id;                                          \
1837                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1838                 cnt++;                                                          \
1839         } while(0);
1840
1841 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1842         do {                                                                    \
1843                 if (FL_KEY_IS_MASKED(mask, member))                             \
1844                         FL_KEY_SET(keys, cnt, id, member);                      \
1845         } while(0);
1846
1847 static void fl_init_dissector(struct flow_dissector *dissector,
1848                               struct fl_flow_key *mask)
1849 {
1850         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1851         size_t cnt = 0;
1852
1853         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1854                              FLOW_DISSECTOR_KEY_META, meta);
1855         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1856         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1857         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1858                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1859         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1860                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1861         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1862                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1863         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1864                              FLOW_DISSECTOR_KEY_PORTS, tp);
1865         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1866                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1867         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1868                              FLOW_DISSECTOR_KEY_IP, ip);
1869         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1870                              FLOW_DISSECTOR_KEY_TCP, tcp);
1871         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1872                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1873         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1874                              FLOW_DISSECTOR_KEY_ARP, arp);
1875         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1876                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1877         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1878                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1879         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1880                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1881         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1882                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1883         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1884                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1885         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1886                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1887         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1888             FL_KEY_IS_MASKED(mask, enc_ipv6))
1889                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1890                            enc_control);
1891         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1892                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1893         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1894                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1895         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1896                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1897         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1898                              FLOW_DISSECTOR_KEY_CT, ct);
1899         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1900                              FLOW_DISSECTOR_KEY_HASH, hash);
1901
1902         skb_flow_dissector_init(dissector, keys, cnt);
1903 }
1904
1905 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1906                                                struct fl_flow_mask *mask)
1907 {
1908         struct fl_flow_mask *newmask;
1909         int err;
1910
1911         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1912         if (!newmask)
1913                 return ERR_PTR(-ENOMEM);
1914
1915         fl_mask_copy(newmask, mask);
1916
1917         if ((newmask->key.tp_range.tp_min.dst &&
1918              newmask->key.tp_range.tp_max.dst) ||
1919             (newmask->key.tp_range.tp_min.src &&
1920              newmask->key.tp_range.tp_max.src))
1921                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1922
1923         err = fl_init_mask_hashtable(newmask);
1924         if (err)
1925                 goto errout_free;
1926
1927         fl_init_dissector(&newmask->dissector, &newmask->key);
1928
1929         INIT_LIST_HEAD_RCU(&newmask->filters);
1930
1931         refcount_set(&newmask->refcnt, 1);
1932         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1933                                       &newmask->ht_node, mask_ht_params);
1934         if (err)
1935                 goto errout_destroy;
1936
1937         spin_lock(&head->masks_lock);
1938         list_add_tail_rcu(&newmask->list, &head->masks);
1939         spin_unlock(&head->masks_lock);
1940
1941         return newmask;
1942
1943 errout_destroy:
1944         rhashtable_destroy(&newmask->ht);
1945 errout_free:
1946         kfree(newmask);
1947
1948         return ERR_PTR(err);
1949 }
1950
1951 static int fl_check_assign_mask(struct cls_fl_head *head,
1952                                 struct cls_fl_filter *fnew,
1953                                 struct cls_fl_filter *fold,
1954                                 struct fl_flow_mask *mask)
1955 {
1956         struct fl_flow_mask *newmask;
1957         int ret = 0;
1958
1959         rcu_read_lock();
1960
1961         /* Insert mask as temporary node to prevent concurrent creation of mask
1962          * with same key. Any concurrent lookups with same key will return
1963          * -EAGAIN because mask's refcnt is zero.
1964          */
1965         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1966                                                        &mask->ht_node,
1967                                                        mask_ht_params);
1968         if (!fnew->mask) {
1969                 rcu_read_unlock();
1970
1971                 if (fold) {
1972                         ret = -EINVAL;
1973                         goto errout_cleanup;
1974                 }
1975
1976                 newmask = fl_create_new_mask(head, mask);
1977                 if (IS_ERR(newmask)) {
1978                         ret = PTR_ERR(newmask);
1979                         goto errout_cleanup;
1980                 }
1981
1982                 fnew->mask = newmask;
1983                 return 0;
1984         } else if (IS_ERR(fnew->mask)) {
1985                 ret = PTR_ERR(fnew->mask);
1986         } else if (fold && fold->mask != fnew->mask) {
1987                 ret = -EINVAL;
1988         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1989                 /* Mask was deleted concurrently, try again */
1990                 ret = -EAGAIN;
1991         }
1992         rcu_read_unlock();
1993         return ret;
1994
1995 errout_cleanup:
1996         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1997                                mask_ht_params);
1998         return ret;
1999 }
2000
2001 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
2002                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
2003                         unsigned long base, struct nlattr **tb,
2004                         struct nlattr *est,
2005                         struct fl_flow_tmplt *tmplt,
2006                         u32 flags, u32 fl_flags,
2007                         struct netlink_ext_ack *extack)
2008 {
2009         int err;
2010
2011         err = tcf_exts_validate_ex(net, tp, tb, est, &f->exts, flags,
2012                                    fl_flags, extack);
2013         if (err < 0)
2014                 return err;
2015
2016         if (tb[TCA_FLOWER_CLASSID]) {
2017                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2018                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2019                         rtnl_lock();
2020                 tcf_bind_filter(tp, &f->res, base);
2021                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2022                         rtnl_unlock();
2023         }
2024
2025         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
2026         if (err)
2027                 return err;
2028
2029         fl_mask_update_range(mask);
2030         fl_set_masked_key(&f->mkey, &f->key, mask);
2031
2032         if (!fl_mask_fits_tmplt(tmplt, mask)) {
2033                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2034                 return -EINVAL;
2035         }
2036
2037         return 0;
2038 }
2039
2040 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2041                                struct cls_fl_filter *fold,
2042                                bool *in_ht)
2043 {
2044         struct fl_flow_mask *mask = fnew->mask;
2045         int err;
2046
2047         err = rhashtable_lookup_insert_fast(&mask->ht,
2048                                             &fnew->ht_node,
2049                                             mask->filter_ht_params);
2050         if (err) {
2051                 *in_ht = false;
2052                 /* It is okay if filter with same key exists when
2053                  * overwriting.
2054                  */
2055                 return fold && err == -EEXIST ? 0 : err;
2056         }
2057
2058         *in_ht = true;
2059         return 0;
2060 }
2061
2062 static int fl_change(struct net *net, struct sk_buff *in_skb,
2063                      struct tcf_proto *tp, unsigned long base,
2064                      u32 handle, struct nlattr **tca,
2065                      void **arg, u32 flags,
2066                      struct netlink_ext_ack *extack)
2067 {
2068         struct cls_fl_head *head = fl_head_dereference(tp);
2069         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2070         struct cls_fl_filter *fold = *arg;
2071         struct cls_fl_filter *fnew;
2072         struct fl_flow_mask *mask;
2073         struct nlattr **tb;
2074         bool in_ht;
2075         int err;
2076
2077         if (!tca[TCA_OPTIONS]) {
2078                 err = -EINVAL;
2079                 goto errout_fold;
2080         }
2081
2082         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2083         if (!mask) {
2084                 err = -ENOBUFS;
2085                 goto errout_fold;
2086         }
2087
2088         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2089         if (!tb) {
2090                 err = -ENOBUFS;
2091                 goto errout_mask_alloc;
2092         }
2093
2094         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2095                                           tca[TCA_OPTIONS], fl_policy, NULL);
2096         if (err < 0)
2097                 goto errout_tb;
2098
2099         if (fold && handle && fold->handle != handle) {
2100                 err = -EINVAL;
2101                 goto errout_tb;
2102         }
2103
2104         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2105         if (!fnew) {
2106                 err = -ENOBUFS;
2107                 goto errout_tb;
2108         }
2109         INIT_LIST_HEAD(&fnew->hw_list);
2110         refcount_set(&fnew->refcnt, 1);
2111
2112         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
2113         if (err < 0)
2114                 goto errout;
2115
2116         if (tb[TCA_FLOWER_FLAGS]) {
2117                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2118
2119                 if (!tc_flags_valid(fnew->flags)) {
2120                         err = -EINVAL;
2121                         goto errout;
2122                 }
2123         }
2124
2125         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2126                            tp->chain->tmplt_priv, flags, fnew->flags,
2127                            extack);
2128         if (err)
2129                 goto errout;
2130
2131         err = fl_check_assign_mask(head, fnew, fold, mask);
2132         if (err)
2133                 goto errout;
2134
2135         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2136         if (err)
2137                 goto errout_mask;
2138
2139         if (!tc_skip_hw(fnew->flags)) {
2140                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2141                 if (err)
2142                         goto errout_ht;
2143         }
2144
2145         if (!tc_in_hw(fnew->flags))
2146                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2147
2148         spin_lock(&tp->lock);
2149
2150         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2151          * proto again or create new one, if necessary.
2152          */
2153         if (tp->deleting) {
2154                 err = -EAGAIN;
2155                 goto errout_hw;
2156         }
2157
2158         if (fold) {
2159                 /* Fold filter was deleted concurrently. Retry lookup. */
2160                 if (fold->deleted) {
2161                         err = -EAGAIN;
2162                         goto errout_hw;
2163                 }
2164
2165                 fnew->handle = handle;
2166
2167                 if (!in_ht) {
2168                         struct rhashtable_params params =
2169                                 fnew->mask->filter_ht_params;
2170
2171                         err = rhashtable_insert_fast(&fnew->mask->ht,
2172                                                      &fnew->ht_node,
2173                                                      params);
2174                         if (err)
2175                                 goto errout_hw;
2176                         in_ht = true;
2177                 }
2178
2179                 refcount_inc(&fnew->refcnt);
2180                 rhashtable_remove_fast(&fold->mask->ht,
2181                                        &fold->ht_node,
2182                                        fold->mask->filter_ht_params);
2183                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2184                 list_replace_rcu(&fold->list, &fnew->list);
2185                 fold->deleted = true;
2186
2187                 spin_unlock(&tp->lock);
2188
2189                 fl_mask_put(head, fold->mask);
2190                 if (!tc_skip_hw(fold->flags))
2191                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2192                 tcf_unbind_filter(tp, &fold->res);
2193                 /* Caller holds reference to fold, so refcnt is always > 0
2194                  * after this.
2195                  */
2196                 refcount_dec(&fold->refcnt);
2197                 __fl_put(fold);
2198         } else {
2199                 if (handle) {
2200                         /* user specifies a handle and it doesn't exist */
2201                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2202                                             handle, GFP_ATOMIC);
2203
2204                         /* Filter with specified handle was concurrently
2205                          * inserted after initial check in cls_api. This is not
2206                          * necessarily an error if NLM_F_EXCL is not set in
2207                          * message flags. Returning EAGAIN will cause cls_api to
2208                          * try to update concurrently inserted rule.
2209                          */
2210                         if (err == -ENOSPC)
2211                                 err = -EAGAIN;
2212                 } else {
2213                         handle = 1;
2214                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2215                                             INT_MAX, GFP_ATOMIC);
2216                 }
2217                 if (err)
2218                         goto errout_hw;
2219
2220                 refcount_inc(&fnew->refcnt);
2221                 fnew->handle = handle;
2222                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2223                 spin_unlock(&tp->lock);
2224         }
2225
2226         *arg = fnew;
2227
2228         kfree(tb);
2229         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2230         return 0;
2231
2232 errout_ht:
2233         spin_lock(&tp->lock);
2234 errout_hw:
2235         fnew->deleted = true;
2236         spin_unlock(&tp->lock);
2237         if (!tc_skip_hw(fnew->flags))
2238                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2239         if (in_ht)
2240                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2241                                        fnew->mask->filter_ht_params);
2242 errout_mask:
2243         fl_mask_put(head, fnew->mask);
2244 errout:
2245         __fl_put(fnew);
2246 errout_tb:
2247         kfree(tb);
2248 errout_mask_alloc:
2249         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2250 errout_fold:
2251         if (fold)
2252                 __fl_put(fold);
2253         return err;
2254 }
2255
2256 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2257                      bool rtnl_held, struct netlink_ext_ack *extack)
2258 {
2259         struct cls_fl_head *head = fl_head_dereference(tp);
2260         struct cls_fl_filter *f = arg;
2261         bool last_on_mask;
2262         int err = 0;
2263
2264         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2265         *last = list_empty(&head->masks);
2266         __fl_put(f);
2267
2268         return err;
2269 }
2270
2271 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2272                     bool rtnl_held)
2273 {
2274         struct cls_fl_head *head = fl_head_dereference(tp);
2275         unsigned long id = arg->cookie, tmp;
2276         struct cls_fl_filter *f;
2277
2278         arg->count = arg->skip;
2279
2280         rcu_read_lock();
2281         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2282                 /* don't return filters that are being deleted */
2283                 if (!refcount_inc_not_zero(&f->refcnt))
2284                         continue;
2285                 rcu_read_unlock();
2286
2287                 if (arg->fn(tp, f, arg) < 0) {
2288                         __fl_put(f);
2289                         arg->stop = 1;
2290                         rcu_read_lock();
2291                         break;
2292                 }
2293                 __fl_put(f);
2294                 arg->count++;
2295                 rcu_read_lock();
2296         }
2297         rcu_read_unlock();
2298         arg->cookie = id;
2299 }
2300
2301 static struct cls_fl_filter *
2302 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2303 {
2304         struct cls_fl_head *head = fl_head_dereference(tp);
2305
2306         spin_lock(&tp->lock);
2307         if (list_empty(&head->hw_filters)) {
2308                 spin_unlock(&tp->lock);
2309                 return NULL;
2310         }
2311
2312         if (!f)
2313                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2314                                hw_list);
2315         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2316                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2317                         spin_unlock(&tp->lock);
2318                         return f;
2319                 }
2320         }
2321
2322         spin_unlock(&tp->lock);
2323         return NULL;
2324 }
2325
2326 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2327                         void *cb_priv, struct netlink_ext_ack *extack)
2328 {
2329         struct tcf_block *block = tp->chain->block;
2330         struct flow_cls_offload cls_flower = {};
2331         struct cls_fl_filter *f = NULL;
2332         int err;
2333
2334         /* hw_filters list can only be changed by hw offload functions after
2335          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2336          * iterating it.
2337          */
2338         ASSERT_RTNL();
2339
2340         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2341                 cls_flower.rule =
2342                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2343                 if (!cls_flower.rule) {
2344                         __fl_put(f);
2345                         return -ENOMEM;
2346                 }
2347
2348                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2349                                            extack);
2350                 cls_flower.command = add ?
2351                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2352                 cls_flower.cookie = (unsigned long)f;
2353                 cls_flower.rule->match.dissector = &f->mask->dissector;
2354                 cls_flower.rule->match.mask = &f->mask->key;
2355                 cls_flower.rule->match.key = &f->mkey;
2356
2357                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts);
2358                 if (err) {
2359                         kfree(cls_flower.rule);
2360                         if (tc_skip_sw(f->flags)) {
2361                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
2362                                 __fl_put(f);
2363                                 return err;
2364                         }
2365                         goto next_flow;
2366                 }
2367
2368                 cls_flower.classid = f->res.classid;
2369
2370                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2371                                             TC_SETUP_CLSFLOWER, &cls_flower,
2372                                             cb_priv, &f->flags,
2373                                             &f->in_hw_count);
2374                 tc_cleanup_offload_action(&cls_flower.rule->action);
2375                 kfree(cls_flower.rule);
2376
2377                 if (err) {
2378                         __fl_put(f);
2379                         return err;
2380                 }
2381 next_flow:
2382                 __fl_put(f);
2383         }
2384
2385         return 0;
2386 }
2387
2388 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2389 {
2390         struct flow_cls_offload *cls_flower = type_data;
2391         struct cls_fl_filter *f =
2392                 (struct cls_fl_filter *) cls_flower->cookie;
2393         struct cls_fl_head *head = fl_head_dereference(tp);
2394
2395         spin_lock(&tp->lock);
2396         list_add(&f->hw_list, &head->hw_filters);
2397         spin_unlock(&tp->lock);
2398 }
2399
2400 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2401 {
2402         struct flow_cls_offload *cls_flower = type_data;
2403         struct cls_fl_filter *f =
2404                 (struct cls_fl_filter *) cls_flower->cookie;
2405
2406         spin_lock(&tp->lock);
2407         if (!list_empty(&f->hw_list))
2408                 list_del_init(&f->hw_list);
2409         spin_unlock(&tp->lock);
2410 }
2411
2412 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2413                               struct fl_flow_tmplt *tmplt)
2414 {
2415         struct flow_cls_offload cls_flower = {};
2416         struct tcf_block *block = chain->block;
2417
2418         cls_flower.rule = flow_rule_alloc(0);
2419         if (!cls_flower.rule)
2420                 return -ENOMEM;
2421
2422         cls_flower.common.chain_index = chain->index;
2423         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2424         cls_flower.cookie = (unsigned long) tmplt;
2425         cls_flower.rule->match.dissector = &tmplt->dissector;
2426         cls_flower.rule->match.mask = &tmplt->mask;
2427         cls_flower.rule->match.key = &tmplt->dummy_key;
2428
2429         /* We don't care if driver (any of them) fails to handle this
2430          * call. It serves just as a hint for it.
2431          */
2432         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2433         kfree(cls_flower.rule);
2434
2435         return 0;
2436 }
2437
2438 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2439                                 struct fl_flow_tmplt *tmplt)
2440 {
2441         struct flow_cls_offload cls_flower = {};
2442         struct tcf_block *block = chain->block;
2443
2444         cls_flower.common.chain_index = chain->index;
2445         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2446         cls_flower.cookie = (unsigned long) tmplt;
2447
2448         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2449 }
2450
2451 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2452                              struct nlattr **tca,
2453                              struct netlink_ext_ack *extack)
2454 {
2455         struct fl_flow_tmplt *tmplt;
2456         struct nlattr **tb;
2457         int err;
2458
2459         if (!tca[TCA_OPTIONS])
2460                 return ERR_PTR(-EINVAL);
2461
2462         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2463         if (!tb)
2464                 return ERR_PTR(-ENOBUFS);
2465         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2466                                           tca[TCA_OPTIONS], fl_policy, NULL);
2467         if (err)
2468                 goto errout_tb;
2469
2470         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2471         if (!tmplt) {
2472                 err = -ENOMEM;
2473                 goto errout_tb;
2474         }
2475         tmplt->chain = chain;
2476         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2477         if (err)
2478                 goto errout_tmplt;
2479
2480         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2481
2482         err = fl_hw_create_tmplt(chain, tmplt);
2483         if (err)
2484                 goto errout_tmplt;
2485
2486         kfree(tb);
2487         return tmplt;
2488
2489 errout_tmplt:
2490         kfree(tmplt);
2491 errout_tb:
2492         kfree(tb);
2493         return ERR_PTR(err);
2494 }
2495
2496 static void fl_tmplt_destroy(void *tmplt_priv)
2497 {
2498         struct fl_flow_tmplt *tmplt = tmplt_priv;
2499
2500         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2501         kfree(tmplt);
2502 }
2503
2504 static int fl_dump_key_val(struct sk_buff *skb,
2505                            void *val, int val_type,
2506                            void *mask, int mask_type, int len)
2507 {
2508         int err;
2509
2510         if (!memchr_inv(mask, 0, len))
2511                 return 0;
2512         err = nla_put(skb, val_type, len, val);
2513         if (err)
2514                 return err;
2515         if (mask_type != TCA_FLOWER_UNSPEC) {
2516                 err = nla_put(skb, mask_type, len, mask);
2517                 if (err)
2518                         return err;
2519         }
2520         return 0;
2521 }
2522
2523 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2524                                   struct fl_flow_key *mask)
2525 {
2526         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2527                             TCA_FLOWER_KEY_PORT_DST_MIN,
2528                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2529                             sizeof(key->tp_range.tp_min.dst)) ||
2530             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2531                             TCA_FLOWER_KEY_PORT_DST_MAX,
2532                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2533                             sizeof(key->tp_range.tp_max.dst)) ||
2534             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2535                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2536                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2537                             sizeof(key->tp_range.tp_min.src)) ||
2538             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2539                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2540                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2541                             sizeof(key->tp_range.tp_max.src)))
2542                 return -1;
2543
2544         return 0;
2545 }
2546
2547 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2548                                     struct flow_dissector_key_mpls *mpls_key,
2549                                     struct flow_dissector_key_mpls *mpls_mask,
2550                                     u8 lse_index)
2551 {
2552         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2553         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2554         int err;
2555
2556         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2557                          lse_index + 1);
2558         if (err)
2559                 return err;
2560
2561         if (lse_mask->mpls_ttl) {
2562                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2563                                  lse_key->mpls_ttl);
2564                 if (err)
2565                         return err;
2566         }
2567         if (lse_mask->mpls_bos) {
2568                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2569                                  lse_key->mpls_bos);
2570                 if (err)
2571                         return err;
2572         }
2573         if (lse_mask->mpls_tc) {
2574                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2575                                  lse_key->mpls_tc);
2576                 if (err)
2577                         return err;
2578         }
2579         if (lse_mask->mpls_label) {
2580                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2581                                   lse_key->mpls_label);
2582                 if (err)
2583                         return err;
2584         }
2585
2586         return 0;
2587 }
2588
2589 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2590                                  struct flow_dissector_key_mpls *mpls_key,
2591                                  struct flow_dissector_key_mpls *mpls_mask)
2592 {
2593         struct nlattr *opts;
2594         struct nlattr *lse;
2595         u8 lse_index;
2596         int err;
2597
2598         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2599         if (!opts)
2600                 return -EMSGSIZE;
2601
2602         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2603                 if (!(mpls_mask->used_lses & 1 << lse_index))
2604                         continue;
2605
2606                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2607                 if (!lse) {
2608                         err = -EMSGSIZE;
2609                         goto err_opts;
2610                 }
2611
2612                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2613                                                lse_index);
2614                 if (err)
2615                         goto err_opts_lse;
2616                 nla_nest_end(skb, lse);
2617         }
2618         nla_nest_end(skb, opts);
2619
2620         return 0;
2621
2622 err_opts_lse:
2623         nla_nest_cancel(skb, lse);
2624 err_opts:
2625         nla_nest_cancel(skb, opts);
2626
2627         return err;
2628 }
2629
2630 static int fl_dump_key_mpls(struct sk_buff *skb,
2631                             struct flow_dissector_key_mpls *mpls_key,
2632                             struct flow_dissector_key_mpls *mpls_mask)
2633 {
2634         struct flow_dissector_mpls_lse *lse_mask;
2635         struct flow_dissector_mpls_lse *lse_key;
2636         int err;
2637
2638         if (!mpls_mask->used_lses)
2639                 return 0;
2640
2641         lse_mask = &mpls_mask->ls[0];
2642         lse_key = &mpls_key->ls[0];
2643
2644         /* For backward compatibility, don't use the MPLS nested attributes if
2645          * the rule can be expressed using the old attributes.
2646          */
2647         if (mpls_mask->used_lses & ~1 ||
2648             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2649              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2650                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2651
2652         if (lse_mask->mpls_ttl) {
2653                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2654                                  lse_key->mpls_ttl);
2655                 if (err)
2656                         return err;
2657         }
2658         if (lse_mask->mpls_tc) {
2659                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2660                                  lse_key->mpls_tc);
2661                 if (err)
2662                         return err;
2663         }
2664         if (lse_mask->mpls_label) {
2665                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2666                                   lse_key->mpls_label);
2667                 if (err)
2668                         return err;
2669         }
2670         if (lse_mask->mpls_bos) {
2671                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2672                                  lse_key->mpls_bos);
2673                 if (err)
2674                         return err;
2675         }
2676         return 0;
2677 }
2678
2679 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2680                           struct flow_dissector_key_ip *key,
2681                           struct flow_dissector_key_ip *mask)
2682 {
2683         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2684         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2685         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2686         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2687
2688         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2689             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2690                 return -1;
2691
2692         return 0;
2693 }
2694
2695 static int fl_dump_key_vlan(struct sk_buff *skb,
2696                             int vlan_id_key, int vlan_prio_key,
2697                             struct flow_dissector_key_vlan *vlan_key,
2698                             struct flow_dissector_key_vlan *vlan_mask)
2699 {
2700         int err;
2701
2702         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2703                 return 0;
2704         if (vlan_mask->vlan_id) {
2705                 err = nla_put_u16(skb, vlan_id_key,
2706                                   vlan_key->vlan_id);
2707                 if (err)
2708                         return err;
2709         }
2710         if (vlan_mask->vlan_priority) {
2711                 err = nla_put_u8(skb, vlan_prio_key,
2712                                  vlan_key->vlan_priority);
2713                 if (err)
2714                         return err;
2715         }
2716         return 0;
2717 }
2718
2719 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2720                             u32 *flower_key, u32 *flower_mask,
2721                             u32 flower_flag_bit, u32 dissector_flag_bit)
2722 {
2723         if (dissector_mask & dissector_flag_bit) {
2724                 *flower_mask |= flower_flag_bit;
2725                 if (dissector_key & dissector_flag_bit)
2726                         *flower_key |= flower_flag_bit;
2727         }
2728 }
2729
2730 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2731 {
2732         u32 key, mask;
2733         __be32 _key, _mask;
2734         int err;
2735
2736         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2737                 return 0;
2738
2739         key = 0;
2740         mask = 0;
2741
2742         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2743                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2744         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2745                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2746                         FLOW_DIS_FIRST_FRAG);
2747
2748         _key = cpu_to_be32(key);
2749         _mask = cpu_to_be32(mask);
2750
2751         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2752         if (err)
2753                 return err;
2754
2755         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2756 }
2757
2758 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2759                                   struct flow_dissector_key_enc_opts *enc_opts)
2760 {
2761         struct geneve_opt *opt;
2762         struct nlattr *nest;
2763         int opt_off = 0;
2764
2765         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2766         if (!nest)
2767                 goto nla_put_failure;
2768
2769         while (enc_opts->len > opt_off) {
2770                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2771
2772                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2773                                  opt->opt_class))
2774                         goto nla_put_failure;
2775                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2776                                opt->type))
2777                         goto nla_put_failure;
2778                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2779                             opt->length * 4, opt->opt_data))
2780                         goto nla_put_failure;
2781
2782                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2783         }
2784         nla_nest_end(skb, nest);
2785         return 0;
2786
2787 nla_put_failure:
2788         nla_nest_cancel(skb, nest);
2789         return -EMSGSIZE;
2790 }
2791
2792 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2793                                  struct flow_dissector_key_enc_opts *enc_opts)
2794 {
2795         struct vxlan_metadata *md;
2796         struct nlattr *nest;
2797
2798         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2799         if (!nest)
2800                 goto nla_put_failure;
2801
2802         md = (struct vxlan_metadata *)&enc_opts->data[0];
2803         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2804                 goto nla_put_failure;
2805
2806         nla_nest_end(skb, nest);
2807         return 0;
2808
2809 nla_put_failure:
2810         nla_nest_cancel(skb, nest);
2811         return -EMSGSIZE;
2812 }
2813
2814 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2815                                   struct flow_dissector_key_enc_opts *enc_opts)
2816 {
2817         struct erspan_metadata *md;
2818         struct nlattr *nest;
2819
2820         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2821         if (!nest)
2822                 goto nla_put_failure;
2823
2824         md = (struct erspan_metadata *)&enc_opts->data[0];
2825         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2826                 goto nla_put_failure;
2827
2828         if (md->version == 1 &&
2829             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2830                 goto nla_put_failure;
2831
2832         if (md->version == 2 &&
2833             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2834                         md->u.md2.dir) ||
2835              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2836                         get_hwid(&md->u.md2))))
2837                 goto nla_put_failure;
2838
2839         nla_nest_end(skb, nest);
2840         return 0;
2841
2842 nla_put_failure:
2843         nla_nest_cancel(skb, nest);
2844         return -EMSGSIZE;
2845 }
2846
2847 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
2848                                struct flow_dissector_key_enc_opts *enc_opts)
2849
2850 {
2851         struct gtp_pdu_session_info *session_info;
2852         struct nlattr *nest;
2853
2854         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
2855         if (!nest)
2856                 goto nla_put_failure;
2857
2858         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
2859
2860         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
2861                        session_info->pdu_type))
2862                 goto nla_put_failure;
2863
2864         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
2865                 goto nla_put_failure;
2866
2867         nla_nest_end(skb, nest);
2868         return 0;
2869
2870 nla_put_failure:
2871         nla_nest_cancel(skb, nest);
2872         return -EMSGSIZE;
2873 }
2874
2875 static int fl_dump_key_ct(struct sk_buff *skb,
2876                           struct flow_dissector_key_ct *key,
2877                           struct flow_dissector_key_ct *mask)
2878 {
2879         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2880             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2881                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2882                             sizeof(key->ct_state)))
2883                 goto nla_put_failure;
2884
2885         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2886             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2887                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2888                             sizeof(key->ct_zone)))
2889                 goto nla_put_failure;
2890
2891         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2892             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2893                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2894                             sizeof(key->ct_mark)))
2895                 goto nla_put_failure;
2896
2897         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2898             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2899                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2900                             sizeof(key->ct_labels)))
2901                 goto nla_put_failure;
2902
2903         return 0;
2904
2905 nla_put_failure:
2906         return -EMSGSIZE;
2907 }
2908
2909 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2910                                struct flow_dissector_key_enc_opts *enc_opts)
2911 {
2912         struct nlattr *nest;
2913         int err;
2914
2915         if (!enc_opts->len)
2916                 return 0;
2917
2918         nest = nla_nest_start_noflag(skb, enc_opt_type);
2919         if (!nest)
2920                 goto nla_put_failure;
2921
2922         switch (enc_opts->dst_opt_type) {
2923         case TUNNEL_GENEVE_OPT:
2924                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2925                 if (err)
2926                         goto nla_put_failure;
2927                 break;
2928         case TUNNEL_VXLAN_OPT:
2929                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
2930                 if (err)
2931                         goto nla_put_failure;
2932                 break;
2933         case TUNNEL_ERSPAN_OPT:
2934                 err = fl_dump_key_erspan_opt(skb, enc_opts);
2935                 if (err)
2936                         goto nla_put_failure;
2937                 break;
2938         case TUNNEL_GTP_OPT:
2939                 err = fl_dump_key_gtp_opt(skb, enc_opts);
2940                 if (err)
2941                         goto nla_put_failure;
2942                 break;
2943         default:
2944                 goto nla_put_failure;
2945         }
2946         nla_nest_end(skb, nest);
2947         return 0;
2948
2949 nla_put_failure:
2950         nla_nest_cancel(skb, nest);
2951         return -EMSGSIZE;
2952 }
2953
2954 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2955                                struct flow_dissector_key_enc_opts *key_opts,
2956                                struct flow_dissector_key_enc_opts *msk_opts)
2957 {
2958         int err;
2959
2960         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2961         if (err)
2962                 return err;
2963
2964         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2965 }
2966
2967 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2968                        struct fl_flow_key *key, struct fl_flow_key *mask)
2969 {
2970         if (mask->meta.ingress_ifindex) {
2971                 struct net_device *dev;
2972
2973                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
2974                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2975                         goto nla_put_failure;
2976         }
2977
2978         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2979                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2980                             sizeof(key->eth.dst)) ||
2981             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2982                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2983                             sizeof(key->eth.src)) ||
2984             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2985                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2986                             sizeof(key->basic.n_proto)))
2987                 goto nla_put_failure;
2988
2989         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2990                 goto nla_put_failure;
2991
2992         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2993                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2994                 goto nla_put_failure;
2995
2996         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2997                              TCA_FLOWER_KEY_CVLAN_PRIO,
2998                              &key->cvlan, &mask->cvlan) ||
2999             (mask->cvlan.vlan_tpid &&
3000              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3001                           key->cvlan.vlan_tpid)))
3002                 goto nla_put_failure;
3003
3004         if (mask->basic.n_proto) {
3005                 if (mask->cvlan.vlan_tpid) {
3006                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3007                                          key->basic.n_proto))
3008                                 goto nla_put_failure;
3009                 } else if (mask->vlan.vlan_tpid) {
3010                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3011                                          key->basic.n_proto))
3012                                 goto nla_put_failure;
3013                 }
3014         }
3015
3016         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3017              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3018             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3019                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3020                             sizeof(key->basic.ip_proto)) ||
3021             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3022                 goto nla_put_failure;
3023
3024         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3025             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3026                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3027                              sizeof(key->ipv4.src)) ||
3028              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3029                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3030                              sizeof(key->ipv4.dst))))
3031                 goto nla_put_failure;
3032         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3033                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3034                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3035                                   sizeof(key->ipv6.src)) ||
3036                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3037                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3038                                   sizeof(key->ipv6.dst))))
3039                 goto nla_put_failure;
3040
3041         if (key->basic.ip_proto == IPPROTO_TCP &&
3042             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3043                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3044                              sizeof(key->tp.src)) ||
3045              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3046                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3047                              sizeof(key->tp.dst)) ||
3048              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3049                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3050                              sizeof(key->tcp.flags))))
3051                 goto nla_put_failure;
3052         else if (key->basic.ip_proto == IPPROTO_UDP &&
3053                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3054                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3055                                   sizeof(key->tp.src)) ||
3056                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3057                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3058                                   sizeof(key->tp.dst))))
3059                 goto nla_put_failure;
3060         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3061                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3062                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3063                                   sizeof(key->tp.src)) ||
3064                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3065                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3066                                   sizeof(key->tp.dst))))
3067                 goto nla_put_failure;
3068         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3069                  key->basic.ip_proto == IPPROTO_ICMP &&
3070                  (fl_dump_key_val(skb, &key->icmp.type,
3071                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3072                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3073                                   sizeof(key->icmp.type)) ||
3074                   fl_dump_key_val(skb, &key->icmp.code,
3075                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3076                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3077                                   sizeof(key->icmp.code))))
3078                 goto nla_put_failure;
3079         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3080                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3081                  (fl_dump_key_val(skb, &key->icmp.type,
3082                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3083                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3084                                   sizeof(key->icmp.type)) ||
3085                   fl_dump_key_val(skb, &key->icmp.code,
3086                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3087                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3088                                   sizeof(key->icmp.code))))
3089                 goto nla_put_failure;
3090         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3091                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3092                  (fl_dump_key_val(skb, &key->arp.sip,
3093                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3094                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3095                                   sizeof(key->arp.sip)) ||
3096                   fl_dump_key_val(skb, &key->arp.tip,
3097                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3098                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3099                                   sizeof(key->arp.tip)) ||
3100                   fl_dump_key_val(skb, &key->arp.op,
3101                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3102                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3103                                   sizeof(key->arp.op)) ||
3104                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3105                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3106                                   sizeof(key->arp.sha)) ||
3107                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3108                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3109                                   sizeof(key->arp.tha))))
3110                 goto nla_put_failure;
3111
3112         if ((key->basic.ip_proto == IPPROTO_TCP ||
3113              key->basic.ip_proto == IPPROTO_UDP ||
3114              key->basic.ip_proto == IPPROTO_SCTP) &&
3115              fl_dump_key_port_range(skb, key, mask))
3116                 goto nla_put_failure;
3117
3118         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3119             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3120                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3121                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3122                             sizeof(key->enc_ipv4.src)) ||
3123              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3124                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3125                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3126                              sizeof(key->enc_ipv4.dst))))
3127                 goto nla_put_failure;
3128         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3129                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3130                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3131                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3132                             sizeof(key->enc_ipv6.src)) ||
3133                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3134                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3135                                  &mask->enc_ipv6.dst,
3136                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3137                             sizeof(key->enc_ipv6.dst))))
3138                 goto nla_put_failure;
3139
3140         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3141                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3142                             sizeof(key->enc_key_id)) ||
3143             fl_dump_key_val(skb, &key->enc_tp.src,
3144                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3145                             &mask->enc_tp.src,
3146                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3147                             sizeof(key->enc_tp.src)) ||
3148             fl_dump_key_val(skb, &key->enc_tp.dst,
3149                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3150                             &mask->enc_tp.dst,
3151                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3152                             sizeof(key->enc_tp.dst)) ||
3153             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3154             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3155                 goto nla_put_failure;
3156
3157         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3158                 goto nla_put_failure;
3159
3160         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3161                 goto nla_put_failure;
3162
3163         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3164                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3165                              sizeof(key->hash.hash)))
3166                 goto nla_put_failure;
3167
3168         return 0;
3169
3170 nla_put_failure:
3171         return -EMSGSIZE;
3172 }
3173
3174 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3175                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3176 {
3177         struct cls_fl_filter *f = fh;
3178         struct nlattr *nest;
3179         struct fl_flow_key *key, *mask;
3180         bool skip_hw;
3181
3182         if (!f)
3183                 return skb->len;
3184
3185         t->tcm_handle = f->handle;
3186
3187         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3188         if (!nest)
3189                 goto nla_put_failure;
3190
3191         spin_lock(&tp->lock);
3192
3193         if (f->res.classid &&
3194             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3195                 goto nla_put_failure_locked;
3196
3197         key = &f->key;
3198         mask = &f->mask->key;
3199         skip_hw = tc_skip_hw(f->flags);
3200
3201         if (fl_dump_key(skb, net, key, mask))
3202                 goto nla_put_failure_locked;
3203
3204         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3205                 goto nla_put_failure_locked;
3206
3207         spin_unlock(&tp->lock);
3208
3209         if (!skip_hw)
3210                 fl_hw_update_stats(tp, f, rtnl_held);
3211
3212         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3213                 goto nla_put_failure;
3214
3215         if (tcf_exts_dump(skb, &f->exts))
3216                 goto nla_put_failure;
3217
3218         nla_nest_end(skb, nest);
3219
3220         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3221                 goto nla_put_failure;
3222
3223         return skb->len;
3224
3225 nla_put_failure_locked:
3226         spin_unlock(&tp->lock);
3227 nla_put_failure:
3228         nla_nest_cancel(skb, nest);
3229         return -1;
3230 }
3231
3232 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3233                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3234 {
3235         struct cls_fl_filter *f = fh;
3236         struct nlattr *nest;
3237         bool skip_hw;
3238
3239         if (!f)
3240                 return skb->len;
3241
3242         t->tcm_handle = f->handle;
3243
3244         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3245         if (!nest)
3246                 goto nla_put_failure;
3247
3248         spin_lock(&tp->lock);
3249
3250         skip_hw = tc_skip_hw(f->flags);
3251
3252         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3253                 goto nla_put_failure_locked;
3254
3255         spin_unlock(&tp->lock);
3256
3257         if (!skip_hw)
3258                 fl_hw_update_stats(tp, f, rtnl_held);
3259
3260         if (tcf_exts_terse_dump(skb, &f->exts))
3261                 goto nla_put_failure;
3262
3263         nla_nest_end(skb, nest);
3264
3265         return skb->len;
3266
3267 nla_put_failure_locked:
3268         spin_unlock(&tp->lock);
3269 nla_put_failure:
3270         nla_nest_cancel(skb, nest);
3271         return -1;
3272 }
3273
3274 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3275 {
3276         struct fl_flow_tmplt *tmplt = tmplt_priv;
3277         struct fl_flow_key *key, *mask;
3278         struct nlattr *nest;
3279
3280         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3281         if (!nest)
3282                 goto nla_put_failure;
3283
3284         key = &tmplt->dummy_key;
3285         mask = &tmplt->mask;
3286
3287         if (fl_dump_key(skb, net, key, mask))
3288                 goto nla_put_failure;
3289
3290         nla_nest_end(skb, nest);
3291
3292         return skb->len;
3293
3294 nla_put_failure:
3295         nla_nest_cancel(skb, nest);
3296         return -EMSGSIZE;
3297 }
3298
3299 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3300                           unsigned long base)
3301 {
3302         struct cls_fl_filter *f = fh;
3303
3304         if (f && f->res.classid == classid) {
3305                 if (cl)
3306                         __tcf_bind_filter(q, &f->res, base);
3307                 else
3308                         __tcf_unbind_filter(q, &f->res);
3309         }
3310 }
3311
3312 static bool fl_delete_empty(struct tcf_proto *tp)
3313 {
3314         struct cls_fl_head *head = fl_head_dereference(tp);
3315
3316         spin_lock(&tp->lock);
3317         tp->deleting = idr_is_empty(&head->handle_idr);
3318         spin_unlock(&tp->lock);
3319
3320         return tp->deleting;
3321 }
3322
3323 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3324         .kind           = "flower",
3325         .classify       = fl_classify,
3326         .init           = fl_init,
3327         .destroy        = fl_destroy,
3328         .get            = fl_get,
3329         .put            = fl_put,
3330         .change         = fl_change,
3331         .delete         = fl_delete,
3332         .delete_empty   = fl_delete_empty,
3333         .walk           = fl_walk,
3334         .reoffload      = fl_reoffload,
3335         .hw_add         = fl_hw_add,
3336         .hw_del         = fl_hw_del,
3337         .dump           = fl_dump,
3338         .terse_dump     = fl_terse_dump,
3339         .bind_class     = fl_bind_class,
3340         .tmplt_create   = fl_tmplt_create,
3341         .tmplt_destroy  = fl_tmplt_destroy,
3342         .tmplt_dump     = fl_tmplt_dump,
3343         .owner          = THIS_MODULE,
3344         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3345 };
3346
3347 static int __init cls_fl_init(void)
3348 {
3349         return register_tcf_proto_ops(&cls_fl_ops);
3350 }
3351
3352 static void __exit cls_fl_exit(void)
3353 {
3354         unregister_tcf_proto_ops(&cls_fl_ops);
3355 }
3356
3357 module_init(cls_fl_init);
3358 module_exit(cls_fl_exit);
3359
3360 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3361 MODULE_DESCRIPTION("Flower classifier");
3362 MODULE_LICENSE("GPL v2");