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                             int vlan_next_eth_type_key,
1017                             struct flow_dissector_key_vlan *key_val,
1018                             struct flow_dissector_key_vlan *key_mask)
1019 {
1020 #define VLAN_PRIORITY_MASK      0x7
1021
1022         if (tb[vlan_id_key]) {
1023                 key_val->vlan_id =
1024                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1025                 key_mask->vlan_id = VLAN_VID_MASK;
1026         }
1027         if (tb[vlan_prio_key]) {
1028                 key_val->vlan_priority =
1029                         nla_get_u8(tb[vlan_prio_key]) &
1030                         VLAN_PRIORITY_MASK;
1031                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1032         }
1033         key_val->vlan_tpid = ethertype;
1034         key_mask->vlan_tpid = cpu_to_be16(~0);
1035         if (tb[vlan_next_eth_type_key]) {
1036                 key_val->vlan_eth_type =
1037                         nla_get_be16(tb[vlan_next_eth_type_key]);
1038                 key_mask->vlan_eth_type = cpu_to_be16(~0);
1039         }
1040 }
1041
1042 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1043                             u32 *dissector_key, u32 *dissector_mask,
1044                             u32 flower_flag_bit, u32 dissector_flag_bit)
1045 {
1046         if (flower_mask & flower_flag_bit) {
1047                 *dissector_mask |= dissector_flag_bit;
1048                 if (flower_key & flower_flag_bit)
1049                         *dissector_key |= dissector_flag_bit;
1050         }
1051 }
1052
1053 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1054                             u32 *flags_mask, struct netlink_ext_ack *extack)
1055 {
1056         u32 key, mask;
1057
1058         /* mask is mandatory for flags */
1059         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1060                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1061                 return -EINVAL;
1062         }
1063
1064         key = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS]));
1065         mask = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1066
1067         *flags_key  = 0;
1068         *flags_mask = 0;
1069
1070         fl_set_key_flag(key, mask, flags_key, flags_mask,
1071                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1072         fl_set_key_flag(key, mask, flags_key, flags_mask,
1073                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1074                         FLOW_DIS_FIRST_FRAG);
1075
1076         return 0;
1077 }
1078
1079 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1080                           struct flow_dissector_key_ip *key,
1081                           struct flow_dissector_key_ip *mask)
1082 {
1083         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1084         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1085         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1086         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1087
1088         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1089         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1090 }
1091
1092 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1093                              int depth, int option_len,
1094                              struct netlink_ext_ack *extack)
1095 {
1096         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1097         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1098         struct geneve_opt *opt;
1099         int err, data_len = 0;
1100
1101         if (option_len > sizeof(struct geneve_opt))
1102                 data_len = option_len - sizeof(struct geneve_opt);
1103
1104         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1105         memset(opt, 0xff, option_len);
1106         opt->length = data_len / 4;
1107         opt->r1 = 0;
1108         opt->r2 = 0;
1109         opt->r3 = 0;
1110
1111         /* If no mask has been prodived we assume an exact match. */
1112         if (!depth)
1113                 return sizeof(struct geneve_opt) + data_len;
1114
1115         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1116                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1117                 return -EINVAL;
1118         }
1119
1120         err = nla_parse_nested_deprecated(tb,
1121                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1122                                           nla, geneve_opt_policy, extack);
1123         if (err < 0)
1124                 return err;
1125
1126         /* We are not allowed to omit any of CLASS, TYPE or DATA
1127          * fields from the key.
1128          */
1129         if (!option_len &&
1130             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1131              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1132              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1133                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1134                 return -EINVAL;
1135         }
1136
1137         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1138          * for the mask.
1139          */
1140         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1141                 int new_len = key->enc_opts.len;
1142
1143                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1144                 data_len = nla_len(data);
1145                 if (data_len < 4) {
1146                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1147                         return -ERANGE;
1148                 }
1149                 if (data_len % 4) {
1150                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1151                         return -ERANGE;
1152                 }
1153
1154                 new_len += sizeof(struct geneve_opt) + data_len;
1155                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1156                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1157                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1158                         return -ERANGE;
1159                 }
1160                 opt->length = data_len / 4;
1161                 memcpy(opt->opt_data, nla_data(data), data_len);
1162         }
1163
1164         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1165                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1166                 opt->opt_class = nla_get_be16(class);
1167         }
1168
1169         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1170                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1171                 opt->type = nla_get_u8(type);
1172         }
1173
1174         return sizeof(struct geneve_opt) + data_len;
1175 }
1176
1177 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1178                             int depth, int option_len,
1179                             struct netlink_ext_ack *extack)
1180 {
1181         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1182         struct vxlan_metadata *md;
1183         int err;
1184
1185         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1186         memset(md, 0xff, sizeof(*md));
1187
1188         if (!depth)
1189                 return sizeof(*md);
1190
1191         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1192                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1193                 return -EINVAL;
1194         }
1195
1196         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1197                                vxlan_opt_policy, extack);
1198         if (err < 0)
1199                 return err;
1200
1201         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1202                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1203                 return -EINVAL;
1204         }
1205
1206         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1207                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1208                 md->gbp &= VXLAN_GBP_MASK;
1209         }
1210
1211         return sizeof(*md);
1212 }
1213
1214 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1215                              int depth, int option_len,
1216                              struct netlink_ext_ack *extack)
1217 {
1218         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1219         struct erspan_metadata *md;
1220         int err;
1221
1222         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1223         memset(md, 0xff, sizeof(*md));
1224         md->version = 1;
1225
1226         if (!depth)
1227                 return sizeof(*md);
1228
1229         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1230                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1231                 return -EINVAL;
1232         }
1233
1234         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1235                                erspan_opt_policy, extack);
1236         if (err < 0)
1237                 return err;
1238
1239         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1240                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1241                 return -EINVAL;
1242         }
1243
1244         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1245                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1246
1247         if (md->version == 1) {
1248                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1249                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1250                         return -EINVAL;
1251                 }
1252                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1253                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1254                         memset(&md->u, 0x00, sizeof(md->u));
1255                         md->u.index = nla_get_be32(nla);
1256                 }
1257         } else if (md->version == 2) {
1258                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1259                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1260                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1261                         return -EINVAL;
1262                 }
1263                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1264                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1265                         md->u.md2.dir = nla_get_u8(nla);
1266                 }
1267                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1268                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1269                         set_hwid(&md->u.md2, nla_get_u8(nla));
1270                 }
1271         } else {
1272                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1273                 return -EINVAL;
1274         }
1275
1276         return sizeof(*md);
1277 }
1278
1279 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1280                           int depth, int option_len,
1281                           struct netlink_ext_ack *extack)
1282 {
1283         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1284         struct gtp_pdu_session_info *sinfo;
1285         u8 len = key->enc_opts.len;
1286         int err;
1287
1288         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1289         memset(sinfo, 0xff, option_len);
1290
1291         if (!depth)
1292                 return sizeof(*sinfo);
1293
1294         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1295                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1296                 return -EINVAL;
1297         }
1298
1299         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1300                                gtp_opt_policy, extack);
1301         if (err < 0)
1302                 return err;
1303
1304         if (!option_len &&
1305             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1306              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1307                 NL_SET_ERR_MSG_MOD(extack,
1308                                    "Missing tunnel key gtp option pdu type or qfi");
1309                 return -EINVAL;
1310         }
1311
1312         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1313                 sinfo->pdu_type =
1314                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1315
1316         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1317                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1318
1319         return sizeof(*sinfo);
1320 }
1321
1322 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1323                           struct fl_flow_key *mask,
1324                           struct netlink_ext_ack *extack)
1325 {
1326         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1327         int err, option_len, key_depth, msk_depth = 0;
1328
1329         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1330                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1331                                              enc_opts_policy, extack);
1332         if (err)
1333                 return err;
1334
1335         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1336
1337         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1338                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1339                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1340                                                      enc_opts_policy, extack);
1341                 if (err)
1342                         return err;
1343
1344                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1345                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1346                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1347                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1348                         return -EINVAL;
1349                 }
1350         }
1351
1352         nla_for_each_attr(nla_opt_key, nla_enc_key,
1353                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1354                 switch (nla_type(nla_opt_key)) {
1355                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1356                         if (key->enc_opts.dst_opt_type &&
1357                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1358                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1359                                 return -EINVAL;
1360                         }
1361                         option_len = 0;
1362                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1363                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1364                                                        key_depth, option_len,
1365                                                        extack);
1366                         if (option_len < 0)
1367                                 return option_len;
1368
1369                         key->enc_opts.len += option_len;
1370                         /* At the same time we need to parse through the mask
1371                          * in order to verify exact and mask attribute lengths.
1372                          */
1373                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1374                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1375                                                        msk_depth, option_len,
1376                                                        extack);
1377                         if (option_len < 0)
1378                                 return option_len;
1379
1380                         mask->enc_opts.len += option_len;
1381                         if (key->enc_opts.len != mask->enc_opts.len) {
1382                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1383                                 return -EINVAL;
1384                         }
1385                         break;
1386                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1387                         if (key->enc_opts.dst_opt_type) {
1388                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1389                                 return -EINVAL;
1390                         }
1391                         option_len = 0;
1392                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1393                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1394                                                       key_depth, option_len,
1395                                                       extack);
1396                         if (option_len < 0)
1397                                 return option_len;
1398
1399                         key->enc_opts.len += option_len;
1400                         /* At the same time we need to parse through the mask
1401                          * in order to verify exact and mask attribute lengths.
1402                          */
1403                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1404                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1405                                                       msk_depth, option_len,
1406                                                       extack);
1407                         if (option_len < 0)
1408                                 return option_len;
1409
1410                         mask->enc_opts.len += option_len;
1411                         if (key->enc_opts.len != mask->enc_opts.len) {
1412                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1413                                 return -EINVAL;
1414                         }
1415                         break;
1416                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1417                         if (key->enc_opts.dst_opt_type) {
1418                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1419                                 return -EINVAL;
1420                         }
1421                         option_len = 0;
1422                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1423                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1424                                                        key_depth, option_len,
1425                                                        extack);
1426                         if (option_len < 0)
1427                                 return option_len;
1428
1429                         key->enc_opts.len += option_len;
1430                         /* At the same time we need to parse through the mask
1431                          * in order to verify exact and mask attribute lengths.
1432                          */
1433                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1434                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1435                                                        msk_depth, option_len,
1436                                                        extack);
1437                         if (option_len < 0)
1438                                 return option_len;
1439
1440                         mask->enc_opts.len += option_len;
1441                         if (key->enc_opts.len != mask->enc_opts.len) {
1442                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1443                                 return -EINVAL;
1444                         }
1445                         break;
1446                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1447                         if (key->enc_opts.dst_opt_type) {
1448                                 NL_SET_ERR_MSG_MOD(extack,
1449                                                    "Duplicate type for gtp options");
1450                                 return -EINVAL;
1451                         }
1452                         option_len = 0;
1453                         key->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1454                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1455                                                     key_depth, option_len,
1456                                                     extack);
1457                         if (option_len < 0)
1458                                 return option_len;
1459
1460                         key->enc_opts.len += option_len;
1461                         /* At the same time we need to parse through the mask
1462                          * in order to verify exact and mask attribute lengths.
1463                          */
1464                         mask->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1465                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1466                                                     msk_depth, option_len,
1467                                                     extack);
1468                         if (option_len < 0)
1469                                 return option_len;
1470
1471                         mask->enc_opts.len += option_len;
1472                         if (key->enc_opts.len != mask->enc_opts.len) {
1473                                 NL_SET_ERR_MSG_MOD(extack,
1474                                                    "Key and mask miss aligned");
1475                                 return -EINVAL;
1476                         }
1477                         break;
1478                 default:
1479                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1480                         return -EINVAL;
1481                 }
1482
1483                 if (!msk_depth)
1484                         continue;
1485
1486                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1487                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1488                         return -EINVAL;
1489                 }
1490                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1491         }
1492
1493         return 0;
1494 }
1495
1496 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1497                                 struct netlink_ext_ack *extack)
1498 {
1499         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1500                 NL_SET_ERR_MSG_ATTR(extack, tb,
1501                                     "no trk, so no other flag can be set");
1502                 return -EINVAL;
1503         }
1504
1505         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1506             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1507                 NL_SET_ERR_MSG_ATTR(extack, tb,
1508                                     "new and est are mutually exclusive");
1509                 return -EINVAL;
1510         }
1511
1512         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1513             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1514                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1515                 NL_SET_ERR_MSG_ATTR(extack, tb,
1516                                     "when inv is set, only trk may be set");
1517                 return -EINVAL;
1518         }
1519
1520         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1521             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1522                 NL_SET_ERR_MSG_ATTR(extack, tb,
1523                                     "new and rpl are mutually exclusive");
1524                 return -EINVAL;
1525         }
1526
1527         return 0;
1528 }
1529
1530 static int fl_set_key_ct(struct nlattr **tb,
1531                          struct flow_dissector_key_ct *key,
1532                          struct flow_dissector_key_ct *mask,
1533                          struct netlink_ext_ack *extack)
1534 {
1535         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1536                 int err;
1537
1538                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1539                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1540                         return -EOPNOTSUPP;
1541                 }
1542                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1543                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1544                                sizeof(key->ct_state));
1545
1546                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1547                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1548                                            extack);
1549                 if (err)
1550                         return err;
1551
1552         }
1553         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1554                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1555                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1556                         return -EOPNOTSUPP;
1557                 }
1558                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1559                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1560                                sizeof(key->ct_zone));
1561         }
1562         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1563                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1564                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1565                         return -EOPNOTSUPP;
1566                 }
1567                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1568                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1569                                sizeof(key->ct_mark));
1570         }
1571         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1572                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1573                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1574                         return -EOPNOTSUPP;
1575                 }
1576                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1577                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1578                                sizeof(key->ct_labels));
1579         }
1580
1581         return 0;
1582 }
1583
1584 static int fl_set_key(struct net *net, struct nlattr **tb,
1585                       struct fl_flow_key *key, struct fl_flow_key *mask,
1586                       struct netlink_ext_ack *extack)
1587 {
1588         __be16 ethertype;
1589         int ret = 0;
1590
1591         if (tb[TCA_FLOWER_INDEV]) {
1592                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1593                 if (err < 0)
1594                         return err;
1595                 key->meta.ingress_ifindex = err;
1596                 mask->meta.ingress_ifindex = 0xffffffff;
1597         }
1598
1599         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1600                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1601                        sizeof(key->eth.dst));
1602         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1603                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1604                        sizeof(key->eth.src));
1605
1606         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1607                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1608
1609                 if (eth_type_vlan(ethertype)) {
1610                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1611                                         TCA_FLOWER_KEY_VLAN_PRIO,
1612                                         TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1613                                         &key->vlan, &mask->vlan);
1614
1615                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1616                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1617                                 if (eth_type_vlan(ethertype)) {
1618                                         fl_set_key_vlan(tb, ethertype,
1619                                                         TCA_FLOWER_KEY_CVLAN_ID,
1620                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1621                                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1622                                                         &key->cvlan, &mask->cvlan);
1623                                         fl_set_key_val(tb, &key->basic.n_proto,
1624                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1625                                                        &mask->basic.n_proto,
1626                                                        TCA_FLOWER_UNSPEC,
1627                                                        sizeof(key->basic.n_proto));
1628                                 } else {
1629                                         key->basic.n_proto = ethertype;
1630                                         mask->basic.n_proto = cpu_to_be16(~0);
1631                                 }
1632                         }
1633                 } else {
1634                         key->basic.n_proto = ethertype;
1635                         mask->basic.n_proto = cpu_to_be16(~0);
1636                 }
1637         }
1638
1639         if (key->basic.n_proto == htons(ETH_P_IP) ||
1640             key->basic.n_proto == htons(ETH_P_IPV6)) {
1641                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1642                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1643                                sizeof(key->basic.ip_proto));
1644                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1645         }
1646
1647         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1648                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1649                 mask->control.addr_type = ~0;
1650                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1651                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1652                                sizeof(key->ipv4.src));
1653                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1654                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1655                                sizeof(key->ipv4.dst));
1656         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1657                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1658                 mask->control.addr_type = ~0;
1659                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1660                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1661                                sizeof(key->ipv6.src));
1662                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1663                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1664                                sizeof(key->ipv6.dst));
1665         }
1666
1667         if (key->basic.ip_proto == IPPROTO_TCP) {
1668                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1669                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1670                                sizeof(key->tp.src));
1671                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1672                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1673                                sizeof(key->tp.dst));
1674                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1675                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1676                                sizeof(key->tcp.flags));
1677         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1678                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1679                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1680                                sizeof(key->tp.src));
1681                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1682                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1683                                sizeof(key->tp.dst));
1684         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1685                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1686                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1687                                sizeof(key->tp.src));
1688                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1689                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1690                                sizeof(key->tp.dst));
1691         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1692                    key->basic.ip_proto == IPPROTO_ICMP) {
1693                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1694                                &mask->icmp.type,
1695                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1696                                sizeof(key->icmp.type));
1697                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1698                                &mask->icmp.code,
1699                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1700                                sizeof(key->icmp.code));
1701         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1702                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1703                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1704                                &mask->icmp.type,
1705                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1706                                sizeof(key->icmp.type));
1707                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1708                                &mask->icmp.code,
1709                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1710                                sizeof(key->icmp.code));
1711         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1712                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1713                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1714                 if (ret)
1715                         return ret;
1716         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1717                    key->basic.n_proto == htons(ETH_P_RARP)) {
1718                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1719                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1720                                sizeof(key->arp.sip));
1721                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1722                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1723                                sizeof(key->arp.tip));
1724                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1725                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1726                                sizeof(key->arp.op));
1727                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1728                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1729                                sizeof(key->arp.sha));
1730                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1731                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1732                                sizeof(key->arp.tha));
1733         }
1734
1735         if (key->basic.ip_proto == IPPROTO_TCP ||
1736             key->basic.ip_proto == IPPROTO_UDP ||
1737             key->basic.ip_proto == IPPROTO_SCTP) {
1738                 ret = fl_set_key_port_range(tb, key, mask, extack);
1739                 if (ret)
1740                         return ret;
1741         }
1742
1743         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1744             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1745                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1746                 mask->enc_control.addr_type = ~0;
1747                 fl_set_key_val(tb, &key->enc_ipv4.src,
1748                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1749                                &mask->enc_ipv4.src,
1750                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1751                                sizeof(key->enc_ipv4.src));
1752                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1753                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1754                                &mask->enc_ipv4.dst,
1755                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1756                                sizeof(key->enc_ipv4.dst));
1757         }
1758
1759         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1760             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1761                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1762                 mask->enc_control.addr_type = ~0;
1763                 fl_set_key_val(tb, &key->enc_ipv6.src,
1764                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1765                                &mask->enc_ipv6.src,
1766                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1767                                sizeof(key->enc_ipv6.src));
1768                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1769                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1770                                &mask->enc_ipv6.dst,
1771                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1772                                sizeof(key->enc_ipv6.dst));
1773         }
1774
1775         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1776                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1777                        sizeof(key->enc_key_id.keyid));
1778
1779         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1780                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1781                        sizeof(key->enc_tp.src));
1782
1783         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1784                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1785                        sizeof(key->enc_tp.dst));
1786
1787         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1788
1789         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1790                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1791                        sizeof(key->hash.hash));
1792
1793         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1794                 ret = fl_set_enc_opt(tb, key, mask, extack);
1795                 if (ret)
1796                         return ret;
1797         }
1798
1799         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1800         if (ret)
1801                 return ret;
1802
1803         if (tb[TCA_FLOWER_KEY_FLAGS])
1804                 ret = fl_set_key_flags(tb, &key->control.flags,
1805                                        &mask->control.flags, extack);
1806
1807         return ret;
1808 }
1809
1810 static void fl_mask_copy(struct fl_flow_mask *dst,
1811                          struct fl_flow_mask *src)
1812 {
1813         const void *psrc = fl_key_get_start(&src->key, src);
1814         void *pdst = fl_key_get_start(&dst->key, src);
1815
1816         memcpy(pdst, psrc, fl_mask_range(src));
1817         dst->range = src->range;
1818 }
1819
1820 static const struct rhashtable_params fl_ht_params = {
1821         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1822         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1823         .automatic_shrinking = true,
1824 };
1825
1826 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1827 {
1828         mask->filter_ht_params = fl_ht_params;
1829         mask->filter_ht_params.key_len = fl_mask_range(mask);
1830         mask->filter_ht_params.key_offset += mask->range.start;
1831
1832         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1833 }
1834
1835 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1836 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1837
1838 #define FL_KEY_IS_MASKED(mask, member)                                          \
1839         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1840                    0, FL_KEY_MEMBER_SIZE(member))                               \
1841
1842 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1843         do {                                                                    \
1844                 keys[cnt].key_id = id;                                          \
1845                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1846                 cnt++;                                                          \
1847         } while(0);
1848
1849 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1850         do {                                                                    \
1851                 if (FL_KEY_IS_MASKED(mask, member))                             \
1852                         FL_KEY_SET(keys, cnt, id, member);                      \
1853         } while(0);
1854
1855 static void fl_init_dissector(struct flow_dissector *dissector,
1856                               struct fl_flow_key *mask)
1857 {
1858         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1859         size_t cnt = 0;
1860
1861         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1862                              FLOW_DISSECTOR_KEY_META, meta);
1863         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1864         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1865         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1866                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1867         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1868                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1869         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1870                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1871         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1872                              FLOW_DISSECTOR_KEY_PORTS, tp);
1873         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1874                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1875         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1876                              FLOW_DISSECTOR_KEY_IP, ip);
1877         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1878                              FLOW_DISSECTOR_KEY_TCP, tcp);
1879         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1880                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1881         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1882                              FLOW_DISSECTOR_KEY_ARP, arp);
1883         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1884                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1885         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1886                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1887         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1888                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1889         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1890                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1891         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1892                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1893         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1894                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1895         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1896             FL_KEY_IS_MASKED(mask, enc_ipv6))
1897                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1898                            enc_control);
1899         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1900                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1901         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1902                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1903         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1904                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1905         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1906                              FLOW_DISSECTOR_KEY_CT, ct);
1907         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1908                              FLOW_DISSECTOR_KEY_HASH, hash);
1909
1910         skb_flow_dissector_init(dissector, keys, cnt);
1911 }
1912
1913 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1914                                                struct fl_flow_mask *mask)
1915 {
1916         struct fl_flow_mask *newmask;
1917         int err;
1918
1919         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1920         if (!newmask)
1921                 return ERR_PTR(-ENOMEM);
1922
1923         fl_mask_copy(newmask, mask);
1924
1925         if ((newmask->key.tp_range.tp_min.dst &&
1926              newmask->key.tp_range.tp_max.dst) ||
1927             (newmask->key.tp_range.tp_min.src &&
1928              newmask->key.tp_range.tp_max.src))
1929                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1930
1931         err = fl_init_mask_hashtable(newmask);
1932         if (err)
1933                 goto errout_free;
1934
1935         fl_init_dissector(&newmask->dissector, &newmask->key);
1936
1937         INIT_LIST_HEAD_RCU(&newmask->filters);
1938
1939         refcount_set(&newmask->refcnt, 1);
1940         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1941                                       &newmask->ht_node, mask_ht_params);
1942         if (err)
1943                 goto errout_destroy;
1944
1945         spin_lock(&head->masks_lock);
1946         list_add_tail_rcu(&newmask->list, &head->masks);
1947         spin_unlock(&head->masks_lock);
1948
1949         return newmask;
1950
1951 errout_destroy:
1952         rhashtable_destroy(&newmask->ht);
1953 errout_free:
1954         kfree(newmask);
1955
1956         return ERR_PTR(err);
1957 }
1958
1959 static int fl_check_assign_mask(struct cls_fl_head *head,
1960                                 struct cls_fl_filter *fnew,
1961                                 struct cls_fl_filter *fold,
1962                                 struct fl_flow_mask *mask)
1963 {
1964         struct fl_flow_mask *newmask;
1965         int ret = 0;
1966
1967         rcu_read_lock();
1968
1969         /* Insert mask as temporary node to prevent concurrent creation of mask
1970          * with same key. Any concurrent lookups with same key will return
1971          * -EAGAIN because mask's refcnt is zero.
1972          */
1973         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1974                                                        &mask->ht_node,
1975                                                        mask_ht_params);
1976         if (!fnew->mask) {
1977                 rcu_read_unlock();
1978
1979                 if (fold) {
1980                         ret = -EINVAL;
1981                         goto errout_cleanup;
1982                 }
1983
1984                 newmask = fl_create_new_mask(head, mask);
1985                 if (IS_ERR(newmask)) {
1986                         ret = PTR_ERR(newmask);
1987                         goto errout_cleanup;
1988                 }
1989
1990                 fnew->mask = newmask;
1991                 return 0;
1992         } else if (IS_ERR(fnew->mask)) {
1993                 ret = PTR_ERR(fnew->mask);
1994         } else if (fold && fold->mask != fnew->mask) {
1995                 ret = -EINVAL;
1996         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1997                 /* Mask was deleted concurrently, try again */
1998                 ret = -EAGAIN;
1999         }
2000         rcu_read_unlock();
2001         return ret;
2002
2003 errout_cleanup:
2004         rhashtable_remove_fast(&head->ht, &mask->ht_node,
2005                                mask_ht_params);
2006         return ret;
2007 }
2008
2009 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
2010                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
2011                         unsigned long base, struct nlattr **tb,
2012                         struct nlattr *est,
2013                         struct fl_flow_tmplt *tmplt,
2014                         u32 flags, u32 fl_flags,
2015                         struct netlink_ext_ack *extack)
2016 {
2017         int err;
2018
2019         err = tcf_exts_validate_ex(net, tp, tb, est, &f->exts, flags,
2020                                    fl_flags, extack);
2021         if (err < 0)
2022                 return err;
2023
2024         if (tb[TCA_FLOWER_CLASSID]) {
2025                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2026                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2027                         rtnl_lock();
2028                 tcf_bind_filter(tp, &f->res, base);
2029                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2030                         rtnl_unlock();
2031         }
2032
2033         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
2034         if (err)
2035                 return err;
2036
2037         fl_mask_update_range(mask);
2038         fl_set_masked_key(&f->mkey, &f->key, mask);
2039
2040         if (!fl_mask_fits_tmplt(tmplt, mask)) {
2041                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2042                 return -EINVAL;
2043         }
2044
2045         return 0;
2046 }
2047
2048 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2049                                struct cls_fl_filter *fold,
2050                                bool *in_ht)
2051 {
2052         struct fl_flow_mask *mask = fnew->mask;
2053         int err;
2054
2055         err = rhashtable_lookup_insert_fast(&mask->ht,
2056                                             &fnew->ht_node,
2057                                             mask->filter_ht_params);
2058         if (err) {
2059                 *in_ht = false;
2060                 /* It is okay if filter with same key exists when
2061                  * overwriting.
2062                  */
2063                 return fold && err == -EEXIST ? 0 : err;
2064         }
2065
2066         *in_ht = true;
2067         return 0;
2068 }
2069
2070 static int fl_change(struct net *net, struct sk_buff *in_skb,
2071                      struct tcf_proto *tp, unsigned long base,
2072                      u32 handle, struct nlattr **tca,
2073                      void **arg, u32 flags,
2074                      struct netlink_ext_ack *extack)
2075 {
2076         struct cls_fl_head *head = fl_head_dereference(tp);
2077         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2078         struct cls_fl_filter *fold = *arg;
2079         struct cls_fl_filter *fnew;
2080         struct fl_flow_mask *mask;
2081         struct nlattr **tb;
2082         bool in_ht;
2083         int err;
2084
2085         if (!tca[TCA_OPTIONS]) {
2086                 err = -EINVAL;
2087                 goto errout_fold;
2088         }
2089
2090         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2091         if (!mask) {
2092                 err = -ENOBUFS;
2093                 goto errout_fold;
2094         }
2095
2096         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2097         if (!tb) {
2098                 err = -ENOBUFS;
2099                 goto errout_mask_alloc;
2100         }
2101
2102         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2103                                           tca[TCA_OPTIONS], fl_policy, NULL);
2104         if (err < 0)
2105                 goto errout_tb;
2106
2107         if (fold && handle && fold->handle != handle) {
2108                 err = -EINVAL;
2109                 goto errout_tb;
2110         }
2111
2112         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2113         if (!fnew) {
2114                 err = -ENOBUFS;
2115                 goto errout_tb;
2116         }
2117         INIT_LIST_HEAD(&fnew->hw_list);
2118         refcount_set(&fnew->refcnt, 1);
2119
2120         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
2121         if (err < 0)
2122                 goto errout;
2123
2124         if (tb[TCA_FLOWER_FLAGS]) {
2125                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2126
2127                 if (!tc_flags_valid(fnew->flags)) {
2128                         err = -EINVAL;
2129                         goto errout;
2130                 }
2131         }
2132
2133         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2134                            tp->chain->tmplt_priv, flags, fnew->flags,
2135                            extack);
2136         if (err)
2137                 goto errout;
2138
2139         err = fl_check_assign_mask(head, fnew, fold, mask);
2140         if (err)
2141                 goto errout;
2142
2143         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2144         if (err)
2145                 goto errout_mask;
2146
2147         if (!tc_skip_hw(fnew->flags)) {
2148                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2149                 if (err)
2150                         goto errout_ht;
2151         }
2152
2153         if (!tc_in_hw(fnew->flags))
2154                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2155
2156         spin_lock(&tp->lock);
2157
2158         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2159          * proto again or create new one, if necessary.
2160          */
2161         if (tp->deleting) {
2162                 err = -EAGAIN;
2163                 goto errout_hw;
2164         }
2165
2166         if (fold) {
2167                 /* Fold filter was deleted concurrently. Retry lookup. */
2168                 if (fold->deleted) {
2169                         err = -EAGAIN;
2170                         goto errout_hw;
2171                 }
2172
2173                 fnew->handle = handle;
2174
2175                 if (!in_ht) {
2176                         struct rhashtable_params params =
2177                                 fnew->mask->filter_ht_params;
2178
2179                         err = rhashtable_insert_fast(&fnew->mask->ht,
2180                                                      &fnew->ht_node,
2181                                                      params);
2182                         if (err)
2183                                 goto errout_hw;
2184                         in_ht = true;
2185                 }
2186
2187                 refcount_inc(&fnew->refcnt);
2188                 rhashtable_remove_fast(&fold->mask->ht,
2189                                        &fold->ht_node,
2190                                        fold->mask->filter_ht_params);
2191                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2192                 list_replace_rcu(&fold->list, &fnew->list);
2193                 fold->deleted = true;
2194
2195                 spin_unlock(&tp->lock);
2196
2197                 fl_mask_put(head, fold->mask);
2198                 if (!tc_skip_hw(fold->flags))
2199                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2200                 tcf_unbind_filter(tp, &fold->res);
2201                 /* Caller holds reference to fold, so refcnt is always > 0
2202                  * after this.
2203                  */
2204                 refcount_dec(&fold->refcnt);
2205                 __fl_put(fold);
2206         } else {
2207                 if (handle) {
2208                         /* user specifies a handle and it doesn't exist */
2209                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2210                                             handle, GFP_ATOMIC);
2211
2212                         /* Filter with specified handle was concurrently
2213                          * inserted after initial check in cls_api. This is not
2214                          * necessarily an error if NLM_F_EXCL is not set in
2215                          * message flags. Returning EAGAIN will cause cls_api to
2216                          * try to update concurrently inserted rule.
2217                          */
2218                         if (err == -ENOSPC)
2219                                 err = -EAGAIN;
2220                 } else {
2221                         handle = 1;
2222                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2223                                             INT_MAX, GFP_ATOMIC);
2224                 }
2225                 if (err)
2226                         goto errout_hw;
2227
2228                 refcount_inc(&fnew->refcnt);
2229                 fnew->handle = handle;
2230                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2231                 spin_unlock(&tp->lock);
2232         }
2233
2234         *arg = fnew;
2235
2236         kfree(tb);
2237         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2238         return 0;
2239
2240 errout_ht:
2241         spin_lock(&tp->lock);
2242 errout_hw:
2243         fnew->deleted = true;
2244         spin_unlock(&tp->lock);
2245         if (!tc_skip_hw(fnew->flags))
2246                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2247         if (in_ht)
2248                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2249                                        fnew->mask->filter_ht_params);
2250 errout_mask:
2251         fl_mask_put(head, fnew->mask);
2252 errout:
2253         __fl_put(fnew);
2254 errout_tb:
2255         kfree(tb);
2256 errout_mask_alloc:
2257         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2258 errout_fold:
2259         if (fold)
2260                 __fl_put(fold);
2261         return err;
2262 }
2263
2264 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2265                      bool rtnl_held, struct netlink_ext_ack *extack)
2266 {
2267         struct cls_fl_head *head = fl_head_dereference(tp);
2268         struct cls_fl_filter *f = arg;
2269         bool last_on_mask;
2270         int err = 0;
2271
2272         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2273         *last = list_empty(&head->masks);
2274         __fl_put(f);
2275
2276         return err;
2277 }
2278
2279 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2280                     bool rtnl_held)
2281 {
2282         struct cls_fl_head *head = fl_head_dereference(tp);
2283         unsigned long id = arg->cookie, tmp;
2284         struct cls_fl_filter *f;
2285
2286         arg->count = arg->skip;
2287
2288         rcu_read_lock();
2289         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2290                 /* don't return filters that are being deleted */
2291                 if (!refcount_inc_not_zero(&f->refcnt))
2292                         continue;
2293                 rcu_read_unlock();
2294
2295                 if (arg->fn(tp, f, arg) < 0) {
2296                         __fl_put(f);
2297                         arg->stop = 1;
2298                         rcu_read_lock();
2299                         break;
2300                 }
2301                 __fl_put(f);
2302                 arg->count++;
2303                 rcu_read_lock();
2304         }
2305         rcu_read_unlock();
2306         arg->cookie = id;
2307 }
2308
2309 static struct cls_fl_filter *
2310 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2311 {
2312         struct cls_fl_head *head = fl_head_dereference(tp);
2313
2314         spin_lock(&tp->lock);
2315         if (list_empty(&head->hw_filters)) {
2316                 spin_unlock(&tp->lock);
2317                 return NULL;
2318         }
2319
2320         if (!f)
2321                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2322                                hw_list);
2323         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2324                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2325                         spin_unlock(&tp->lock);
2326                         return f;
2327                 }
2328         }
2329
2330         spin_unlock(&tp->lock);
2331         return NULL;
2332 }
2333
2334 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2335                         void *cb_priv, struct netlink_ext_ack *extack)
2336 {
2337         struct tcf_block *block = tp->chain->block;
2338         struct flow_cls_offload cls_flower = {};
2339         struct cls_fl_filter *f = NULL;
2340         int err;
2341
2342         /* hw_filters list can only be changed by hw offload functions after
2343          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2344          * iterating it.
2345          */
2346         ASSERT_RTNL();
2347
2348         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2349                 cls_flower.rule =
2350                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2351                 if (!cls_flower.rule) {
2352                         __fl_put(f);
2353                         return -ENOMEM;
2354                 }
2355
2356                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2357                                            extack);
2358                 cls_flower.command = add ?
2359                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2360                 cls_flower.cookie = (unsigned long)f;
2361                 cls_flower.rule->match.dissector = &f->mask->dissector;
2362                 cls_flower.rule->match.mask = &f->mask->key;
2363                 cls_flower.rule->match.key = &f->mkey;
2364
2365                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts);
2366                 if (err) {
2367                         kfree(cls_flower.rule);
2368                         if (tc_skip_sw(f->flags)) {
2369                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
2370                                 __fl_put(f);
2371                                 return err;
2372                         }
2373                         goto next_flow;
2374                 }
2375
2376                 cls_flower.classid = f->res.classid;
2377
2378                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2379                                             TC_SETUP_CLSFLOWER, &cls_flower,
2380                                             cb_priv, &f->flags,
2381                                             &f->in_hw_count);
2382                 tc_cleanup_offload_action(&cls_flower.rule->action);
2383                 kfree(cls_flower.rule);
2384
2385                 if (err) {
2386                         __fl_put(f);
2387                         return err;
2388                 }
2389 next_flow:
2390                 __fl_put(f);
2391         }
2392
2393         return 0;
2394 }
2395
2396 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2397 {
2398         struct flow_cls_offload *cls_flower = type_data;
2399         struct cls_fl_filter *f =
2400                 (struct cls_fl_filter *) cls_flower->cookie;
2401         struct cls_fl_head *head = fl_head_dereference(tp);
2402
2403         spin_lock(&tp->lock);
2404         list_add(&f->hw_list, &head->hw_filters);
2405         spin_unlock(&tp->lock);
2406 }
2407
2408 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2409 {
2410         struct flow_cls_offload *cls_flower = type_data;
2411         struct cls_fl_filter *f =
2412                 (struct cls_fl_filter *) cls_flower->cookie;
2413
2414         spin_lock(&tp->lock);
2415         if (!list_empty(&f->hw_list))
2416                 list_del_init(&f->hw_list);
2417         spin_unlock(&tp->lock);
2418 }
2419
2420 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2421                               struct fl_flow_tmplt *tmplt)
2422 {
2423         struct flow_cls_offload cls_flower = {};
2424         struct tcf_block *block = chain->block;
2425
2426         cls_flower.rule = flow_rule_alloc(0);
2427         if (!cls_flower.rule)
2428                 return -ENOMEM;
2429
2430         cls_flower.common.chain_index = chain->index;
2431         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2432         cls_flower.cookie = (unsigned long) tmplt;
2433         cls_flower.rule->match.dissector = &tmplt->dissector;
2434         cls_flower.rule->match.mask = &tmplt->mask;
2435         cls_flower.rule->match.key = &tmplt->dummy_key;
2436
2437         /* We don't care if driver (any of them) fails to handle this
2438          * call. It serves just as a hint for it.
2439          */
2440         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2441         kfree(cls_flower.rule);
2442
2443         return 0;
2444 }
2445
2446 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2447                                 struct fl_flow_tmplt *tmplt)
2448 {
2449         struct flow_cls_offload cls_flower = {};
2450         struct tcf_block *block = chain->block;
2451
2452         cls_flower.common.chain_index = chain->index;
2453         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2454         cls_flower.cookie = (unsigned long) tmplt;
2455
2456         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2457 }
2458
2459 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2460                              struct nlattr **tca,
2461                              struct netlink_ext_ack *extack)
2462 {
2463         struct fl_flow_tmplt *tmplt;
2464         struct nlattr **tb;
2465         int err;
2466
2467         if (!tca[TCA_OPTIONS])
2468                 return ERR_PTR(-EINVAL);
2469
2470         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2471         if (!tb)
2472                 return ERR_PTR(-ENOBUFS);
2473         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2474                                           tca[TCA_OPTIONS], fl_policy, NULL);
2475         if (err)
2476                 goto errout_tb;
2477
2478         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2479         if (!tmplt) {
2480                 err = -ENOMEM;
2481                 goto errout_tb;
2482         }
2483         tmplt->chain = chain;
2484         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2485         if (err)
2486                 goto errout_tmplt;
2487
2488         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2489
2490         err = fl_hw_create_tmplt(chain, tmplt);
2491         if (err)
2492                 goto errout_tmplt;
2493
2494         kfree(tb);
2495         return tmplt;
2496
2497 errout_tmplt:
2498         kfree(tmplt);
2499 errout_tb:
2500         kfree(tb);
2501         return ERR_PTR(err);
2502 }
2503
2504 static void fl_tmplt_destroy(void *tmplt_priv)
2505 {
2506         struct fl_flow_tmplt *tmplt = tmplt_priv;
2507
2508         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2509         kfree(tmplt);
2510 }
2511
2512 static int fl_dump_key_val(struct sk_buff *skb,
2513                            void *val, int val_type,
2514                            void *mask, int mask_type, int len)
2515 {
2516         int err;
2517
2518         if (!memchr_inv(mask, 0, len))
2519                 return 0;
2520         err = nla_put(skb, val_type, len, val);
2521         if (err)
2522                 return err;
2523         if (mask_type != TCA_FLOWER_UNSPEC) {
2524                 err = nla_put(skb, mask_type, len, mask);
2525                 if (err)
2526                         return err;
2527         }
2528         return 0;
2529 }
2530
2531 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2532                                   struct fl_flow_key *mask)
2533 {
2534         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2535                             TCA_FLOWER_KEY_PORT_DST_MIN,
2536                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2537                             sizeof(key->tp_range.tp_min.dst)) ||
2538             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2539                             TCA_FLOWER_KEY_PORT_DST_MAX,
2540                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2541                             sizeof(key->tp_range.tp_max.dst)) ||
2542             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2543                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2544                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2545                             sizeof(key->tp_range.tp_min.src)) ||
2546             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2547                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2548                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2549                             sizeof(key->tp_range.tp_max.src)))
2550                 return -1;
2551
2552         return 0;
2553 }
2554
2555 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2556                                     struct flow_dissector_key_mpls *mpls_key,
2557                                     struct flow_dissector_key_mpls *mpls_mask,
2558                                     u8 lse_index)
2559 {
2560         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2561         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2562         int err;
2563
2564         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2565                          lse_index + 1);
2566         if (err)
2567                 return err;
2568
2569         if (lse_mask->mpls_ttl) {
2570                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2571                                  lse_key->mpls_ttl);
2572                 if (err)
2573                         return err;
2574         }
2575         if (lse_mask->mpls_bos) {
2576                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2577                                  lse_key->mpls_bos);
2578                 if (err)
2579                         return err;
2580         }
2581         if (lse_mask->mpls_tc) {
2582                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2583                                  lse_key->mpls_tc);
2584                 if (err)
2585                         return err;
2586         }
2587         if (lse_mask->mpls_label) {
2588                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2589                                   lse_key->mpls_label);
2590                 if (err)
2591                         return err;
2592         }
2593
2594         return 0;
2595 }
2596
2597 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2598                                  struct flow_dissector_key_mpls *mpls_key,
2599                                  struct flow_dissector_key_mpls *mpls_mask)
2600 {
2601         struct nlattr *opts;
2602         struct nlattr *lse;
2603         u8 lse_index;
2604         int err;
2605
2606         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2607         if (!opts)
2608                 return -EMSGSIZE;
2609
2610         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2611                 if (!(mpls_mask->used_lses & 1 << lse_index))
2612                         continue;
2613
2614                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2615                 if (!lse) {
2616                         err = -EMSGSIZE;
2617                         goto err_opts;
2618                 }
2619
2620                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2621                                                lse_index);
2622                 if (err)
2623                         goto err_opts_lse;
2624                 nla_nest_end(skb, lse);
2625         }
2626         nla_nest_end(skb, opts);
2627
2628         return 0;
2629
2630 err_opts_lse:
2631         nla_nest_cancel(skb, lse);
2632 err_opts:
2633         nla_nest_cancel(skb, opts);
2634
2635         return err;
2636 }
2637
2638 static int fl_dump_key_mpls(struct sk_buff *skb,
2639                             struct flow_dissector_key_mpls *mpls_key,
2640                             struct flow_dissector_key_mpls *mpls_mask)
2641 {
2642         struct flow_dissector_mpls_lse *lse_mask;
2643         struct flow_dissector_mpls_lse *lse_key;
2644         int err;
2645
2646         if (!mpls_mask->used_lses)
2647                 return 0;
2648
2649         lse_mask = &mpls_mask->ls[0];
2650         lse_key = &mpls_key->ls[0];
2651
2652         /* For backward compatibility, don't use the MPLS nested attributes if
2653          * the rule can be expressed using the old attributes.
2654          */
2655         if (mpls_mask->used_lses & ~1 ||
2656             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2657              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2658                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2659
2660         if (lse_mask->mpls_ttl) {
2661                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2662                                  lse_key->mpls_ttl);
2663                 if (err)
2664                         return err;
2665         }
2666         if (lse_mask->mpls_tc) {
2667                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2668                                  lse_key->mpls_tc);
2669                 if (err)
2670                         return err;
2671         }
2672         if (lse_mask->mpls_label) {
2673                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2674                                   lse_key->mpls_label);
2675                 if (err)
2676                         return err;
2677         }
2678         if (lse_mask->mpls_bos) {
2679                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2680                                  lse_key->mpls_bos);
2681                 if (err)
2682                         return err;
2683         }
2684         return 0;
2685 }
2686
2687 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2688                           struct flow_dissector_key_ip *key,
2689                           struct flow_dissector_key_ip *mask)
2690 {
2691         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2692         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2693         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2694         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2695
2696         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2697             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2698                 return -1;
2699
2700         return 0;
2701 }
2702
2703 static int fl_dump_key_vlan(struct sk_buff *skb,
2704                             int vlan_id_key, int vlan_prio_key,
2705                             struct flow_dissector_key_vlan *vlan_key,
2706                             struct flow_dissector_key_vlan *vlan_mask)
2707 {
2708         int err;
2709
2710         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2711                 return 0;
2712         if (vlan_mask->vlan_id) {
2713                 err = nla_put_u16(skb, vlan_id_key,
2714                                   vlan_key->vlan_id);
2715                 if (err)
2716                         return err;
2717         }
2718         if (vlan_mask->vlan_priority) {
2719                 err = nla_put_u8(skb, vlan_prio_key,
2720                                  vlan_key->vlan_priority);
2721                 if (err)
2722                         return err;
2723         }
2724         return 0;
2725 }
2726
2727 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2728                             u32 *flower_key, u32 *flower_mask,
2729                             u32 flower_flag_bit, u32 dissector_flag_bit)
2730 {
2731         if (dissector_mask & dissector_flag_bit) {
2732                 *flower_mask |= flower_flag_bit;
2733                 if (dissector_key & dissector_flag_bit)
2734                         *flower_key |= flower_flag_bit;
2735         }
2736 }
2737
2738 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2739 {
2740         u32 key, mask;
2741         __be32 _key, _mask;
2742         int err;
2743
2744         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2745                 return 0;
2746
2747         key = 0;
2748         mask = 0;
2749
2750         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2751                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2752         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2753                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2754                         FLOW_DIS_FIRST_FRAG);
2755
2756         _key = cpu_to_be32(key);
2757         _mask = cpu_to_be32(mask);
2758
2759         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2760         if (err)
2761                 return err;
2762
2763         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2764 }
2765
2766 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2767                                   struct flow_dissector_key_enc_opts *enc_opts)
2768 {
2769         struct geneve_opt *opt;
2770         struct nlattr *nest;
2771         int opt_off = 0;
2772
2773         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2774         if (!nest)
2775                 goto nla_put_failure;
2776
2777         while (enc_opts->len > opt_off) {
2778                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2779
2780                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2781                                  opt->opt_class))
2782                         goto nla_put_failure;
2783                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2784                                opt->type))
2785                         goto nla_put_failure;
2786                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2787                             opt->length * 4, opt->opt_data))
2788                         goto nla_put_failure;
2789
2790                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2791         }
2792         nla_nest_end(skb, nest);
2793         return 0;
2794
2795 nla_put_failure:
2796         nla_nest_cancel(skb, nest);
2797         return -EMSGSIZE;
2798 }
2799
2800 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2801                                  struct flow_dissector_key_enc_opts *enc_opts)
2802 {
2803         struct vxlan_metadata *md;
2804         struct nlattr *nest;
2805
2806         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2807         if (!nest)
2808                 goto nla_put_failure;
2809
2810         md = (struct vxlan_metadata *)&enc_opts->data[0];
2811         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2812                 goto nla_put_failure;
2813
2814         nla_nest_end(skb, nest);
2815         return 0;
2816
2817 nla_put_failure:
2818         nla_nest_cancel(skb, nest);
2819         return -EMSGSIZE;
2820 }
2821
2822 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2823                                   struct flow_dissector_key_enc_opts *enc_opts)
2824 {
2825         struct erspan_metadata *md;
2826         struct nlattr *nest;
2827
2828         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2829         if (!nest)
2830                 goto nla_put_failure;
2831
2832         md = (struct erspan_metadata *)&enc_opts->data[0];
2833         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2834                 goto nla_put_failure;
2835
2836         if (md->version == 1 &&
2837             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2838                 goto nla_put_failure;
2839
2840         if (md->version == 2 &&
2841             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2842                         md->u.md2.dir) ||
2843              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2844                         get_hwid(&md->u.md2))))
2845                 goto nla_put_failure;
2846
2847         nla_nest_end(skb, nest);
2848         return 0;
2849
2850 nla_put_failure:
2851         nla_nest_cancel(skb, nest);
2852         return -EMSGSIZE;
2853 }
2854
2855 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
2856                                struct flow_dissector_key_enc_opts *enc_opts)
2857
2858 {
2859         struct gtp_pdu_session_info *session_info;
2860         struct nlattr *nest;
2861
2862         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
2863         if (!nest)
2864                 goto nla_put_failure;
2865
2866         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
2867
2868         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
2869                        session_info->pdu_type))
2870                 goto nla_put_failure;
2871
2872         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
2873                 goto nla_put_failure;
2874
2875         nla_nest_end(skb, nest);
2876         return 0;
2877
2878 nla_put_failure:
2879         nla_nest_cancel(skb, nest);
2880         return -EMSGSIZE;
2881 }
2882
2883 static int fl_dump_key_ct(struct sk_buff *skb,
2884                           struct flow_dissector_key_ct *key,
2885                           struct flow_dissector_key_ct *mask)
2886 {
2887         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2888             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2889                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2890                             sizeof(key->ct_state)))
2891                 goto nla_put_failure;
2892
2893         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2894             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2895                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2896                             sizeof(key->ct_zone)))
2897                 goto nla_put_failure;
2898
2899         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2900             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2901                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2902                             sizeof(key->ct_mark)))
2903                 goto nla_put_failure;
2904
2905         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2906             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2907                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2908                             sizeof(key->ct_labels)))
2909                 goto nla_put_failure;
2910
2911         return 0;
2912
2913 nla_put_failure:
2914         return -EMSGSIZE;
2915 }
2916
2917 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2918                                struct flow_dissector_key_enc_opts *enc_opts)
2919 {
2920         struct nlattr *nest;
2921         int err;
2922
2923         if (!enc_opts->len)
2924                 return 0;
2925
2926         nest = nla_nest_start_noflag(skb, enc_opt_type);
2927         if (!nest)
2928                 goto nla_put_failure;
2929
2930         switch (enc_opts->dst_opt_type) {
2931         case TUNNEL_GENEVE_OPT:
2932                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2933                 if (err)
2934                         goto nla_put_failure;
2935                 break;
2936         case TUNNEL_VXLAN_OPT:
2937                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
2938                 if (err)
2939                         goto nla_put_failure;
2940                 break;
2941         case TUNNEL_ERSPAN_OPT:
2942                 err = fl_dump_key_erspan_opt(skb, enc_opts);
2943                 if (err)
2944                         goto nla_put_failure;
2945                 break;
2946         case TUNNEL_GTP_OPT:
2947                 err = fl_dump_key_gtp_opt(skb, enc_opts);
2948                 if (err)
2949                         goto nla_put_failure;
2950                 break;
2951         default:
2952                 goto nla_put_failure;
2953         }
2954         nla_nest_end(skb, nest);
2955         return 0;
2956
2957 nla_put_failure:
2958         nla_nest_cancel(skb, nest);
2959         return -EMSGSIZE;
2960 }
2961
2962 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2963                                struct flow_dissector_key_enc_opts *key_opts,
2964                                struct flow_dissector_key_enc_opts *msk_opts)
2965 {
2966         int err;
2967
2968         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2969         if (err)
2970                 return err;
2971
2972         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2973 }
2974
2975 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2976                        struct fl_flow_key *key, struct fl_flow_key *mask)
2977 {
2978         if (mask->meta.ingress_ifindex) {
2979                 struct net_device *dev;
2980
2981                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
2982                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2983                         goto nla_put_failure;
2984         }
2985
2986         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2987                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2988                             sizeof(key->eth.dst)) ||
2989             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2990                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2991                             sizeof(key->eth.src)) ||
2992             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2993                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2994                             sizeof(key->basic.n_proto)))
2995                 goto nla_put_failure;
2996
2997         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2998                 goto nla_put_failure;
2999
3000         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
3001                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
3002                 goto nla_put_failure;
3003
3004         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
3005                              TCA_FLOWER_KEY_CVLAN_PRIO,
3006                              &key->cvlan, &mask->cvlan) ||
3007             (mask->cvlan.vlan_tpid &&
3008              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3009                           key->cvlan.vlan_tpid)))
3010                 goto nla_put_failure;
3011
3012         if (mask->basic.n_proto) {
3013                 if (mask->cvlan.vlan_eth_type) {
3014                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3015                                          key->basic.n_proto))
3016                                 goto nla_put_failure;
3017                 } else if (mask->vlan.vlan_eth_type) {
3018                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3019                                          key->vlan.vlan_eth_type))
3020                                 goto nla_put_failure;
3021                 }
3022         }
3023
3024         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3025              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3026             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3027                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3028                             sizeof(key->basic.ip_proto)) ||
3029             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3030                 goto nla_put_failure;
3031
3032         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3033             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3034                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3035                              sizeof(key->ipv4.src)) ||
3036              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3037                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3038                              sizeof(key->ipv4.dst))))
3039                 goto nla_put_failure;
3040         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3041                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3042                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3043                                   sizeof(key->ipv6.src)) ||
3044                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3045                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3046                                   sizeof(key->ipv6.dst))))
3047                 goto nla_put_failure;
3048
3049         if (key->basic.ip_proto == IPPROTO_TCP &&
3050             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3051                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3052                              sizeof(key->tp.src)) ||
3053              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3054                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3055                              sizeof(key->tp.dst)) ||
3056              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3057                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3058                              sizeof(key->tcp.flags))))
3059                 goto nla_put_failure;
3060         else if (key->basic.ip_proto == IPPROTO_UDP &&
3061                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3062                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3063                                   sizeof(key->tp.src)) ||
3064                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3065                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3066                                   sizeof(key->tp.dst))))
3067                 goto nla_put_failure;
3068         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3069                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3070                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3071                                   sizeof(key->tp.src)) ||
3072                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3073                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3074                                   sizeof(key->tp.dst))))
3075                 goto nla_put_failure;
3076         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3077                  key->basic.ip_proto == IPPROTO_ICMP &&
3078                  (fl_dump_key_val(skb, &key->icmp.type,
3079                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3080                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3081                                   sizeof(key->icmp.type)) ||
3082                   fl_dump_key_val(skb, &key->icmp.code,
3083                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3084                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3085                                   sizeof(key->icmp.code))))
3086                 goto nla_put_failure;
3087         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3088                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3089                  (fl_dump_key_val(skb, &key->icmp.type,
3090                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3091                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3092                                   sizeof(key->icmp.type)) ||
3093                   fl_dump_key_val(skb, &key->icmp.code,
3094                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3095                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3096                                   sizeof(key->icmp.code))))
3097                 goto nla_put_failure;
3098         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3099                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3100                  (fl_dump_key_val(skb, &key->arp.sip,
3101                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3102                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3103                                   sizeof(key->arp.sip)) ||
3104                   fl_dump_key_val(skb, &key->arp.tip,
3105                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3106                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3107                                   sizeof(key->arp.tip)) ||
3108                   fl_dump_key_val(skb, &key->arp.op,
3109                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3110                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3111                                   sizeof(key->arp.op)) ||
3112                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3113                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3114                                   sizeof(key->arp.sha)) ||
3115                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3116                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3117                                   sizeof(key->arp.tha))))
3118                 goto nla_put_failure;
3119
3120         if ((key->basic.ip_proto == IPPROTO_TCP ||
3121              key->basic.ip_proto == IPPROTO_UDP ||
3122              key->basic.ip_proto == IPPROTO_SCTP) &&
3123              fl_dump_key_port_range(skb, key, mask))
3124                 goto nla_put_failure;
3125
3126         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3127             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3128                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3129                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3130                             sizeof(key->enc_ipv4.src)) ||
3131              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3132                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3133                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3134                              sizeof(key->enc_ipv4.dst))))
3135                 goto nla_put_failure;
3136         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3137                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3138                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3139                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3140                             sizeof(key->enc_ipv6.src)) ||
3141                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3142                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3143                                  &mask->enc_ipv6.dst,
3144                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3145                             sizeof(key->enc_ipv6.dst))))
3146                 goto nla_put_failure;
3147
3148         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3149                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3150                             sizeof(key->enc_key_id)) ||
3151             fl_dump_key_val(skb, &key->enc_tp.src,
3152                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3153                             &mask->enc_tp.src,
3154                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3155                             sizeof(key->enc_tp.src)) ||
3156             fl_dump_key_val(skb, &key->enc_tp.dst,
3157                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3158                             &mask->enc_tp.dst,
3159                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3160                             sizeof(key->enc_tp.dst)) ||
3161             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3162             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3163                 goto nla_put_failure;
3164
3165         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3166                 goto nla_put_failure;
3167
3168         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3169                 goto nla_put_failure;
3170
3171         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3172                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3173                              sizeof(key->hash.hash)))
3174                 goto nla_put_failure;
3175
3176         return 0;
3177
3178 nla_put_failure:
3179         return -EMSGSIZE;
3180 }
3181
3182 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3183                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3184 {
3185         struct cls_fl_filter *f = fh;
3186         struct nlattr *nest;
3187         struct fl_flow_key *key, *mask;
3188         bool skip_hw;
3189
3190         if (!f)
3191                 return skb->len;
3192
3193         t->tcm_handle = f->handle;
3194
3195         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3196         if (!nest)
3197                 goto nla_put_failure;
3198
3199         spin_lock(&tp->lock);
3200
3201         if (f->res.classid &&
3202             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3203                 goto nla_put_failure_locked;
3204
3205         key = &f->key;
3206         mask = &f->mask->key;
3207         skip_hw = tc_skip_hw(f->flags);
3208
3209         if (fl_dump_key(skb, net, key, mask))
3210                 goto nla_put_failure_locked;
3211
3212         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3213                 goto nla_put_failure_locked;
3214
3215         spin_unlock(&tp->lock);
3216
3217         if (!skip_hw)
3218                 fl_hw_update_stats(tp, f, rtnl_held);
3219
3220         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3221                 goto nla_put_failure;
3222
3223         if (tcf_exts_dump(skb, &f->exts))
3224                 goto nla_put_failure;
3225
3226         nla_nest_end(skb, nest);
3227
3228         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3229                 goto nla_put_failure;
3230
3231         return skb->len;
3232
3233 nla_put_failure_locked:
3234         spin_unlock(&tp->lock);
3235 nla_put_failure:
3236         nla_nest_cancel(skb, nest);
3237         return -1;
3238 }
3239
3240 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3241                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3242 {
3243         struct cls_fl_filter *f = fh;
3244         struct nlattr *nest;
3245         bool skip_hw;
3246
3247         if (!f)
3248                 return skb->len;
3249
3250         t->tcm_handle = f->handle;
3251
3252         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3253         if (!nest)
3254                 goto nla_put_failure;
3255
3256         spin_lock(&tp->lock);
3257
3258         skip_hw = tc_skip_hw(f->flags);
3259
3260         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3261                 goto nla_put_failure_locked;
3262
3263         spin_unlock(&tp->lock);
3264
3265         if (!skip_hw)
3266                 fl_hw_update_stats(tp, f, rtnl_held);
3267
3268         if (tcf_exts_terse_dump(skb, &f->exts))
3269                 goto nla_put_failure;
3270
3271         nla_nest_end(skb, nest);
3272
3273         return skb->len;
3274
3275 nla_put_failure_locked:
3276         spin_unlock(&tp->lock);
3277 nla_put_failure:
3278         nla_nest_cancel(skb, nest);
3279         return -1;
3280 }
3281
3282 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3283 {
3284         struct fl_flow_tmplt *tmplt = tmplt_priv;
3285         struct fl_flow_key *key, *mask;
3286         struct nlattr *nest;
3287
3288         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3289         if (!nest)
3290                 goto nla_put_failure;
3291
3292         key = &tmplt->dummy_key;
3293         mask = &tmplt->mask;
3294
3295         if (fl_dump_key(skb, net, key, mask))
3296                 goto nla_put_failure;
3297
3298         nla_nest_end(skb, nest);
3299
3300         return skb->len;
3301
3302 nla_put_failure:
3303         nla_nest_cancel(skb, nest);
3304         return -EMSGSIZE;
3305 }
3306
3307 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3308                           unsigned long base)
3309 {
3310         struct cls_fl_filter *f = fh;
3311
3312         if (f && f->res.classid == classid) {
3313                 if (cl)
3314                         __tcf_bind_filter(q, &f->res, base);
3315                 else
3316                         __tcf_unbind_filter(q, &f->res);
3317         }
3318 }
3319
3320 static bool fl_delete_empty(struct tcf_proto *tp)
3321 {
3322         struct cls_fl_head *head = fl_head_dereference(tp);
3323
3324         spin_lock(&tp->lock);
3325         tp->deleting = idr_is_empty(&head->handle_idr);
3326         spin_unlock(&tp->lock);
3327
3328         return tp->deleting;
3329 }
3330
3331 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3332         .kind           = "flower",
3333         .classify       = fl_classify,
3334         .init           = fl_init,
3335         .destroy        = fl_destroy,
3336         .get            = fl_get,
3337         .put            = fl_put,
3338         .change         = fl_change,
3339         .delete         = fl_delete,
3340         .delete_empty   = fl_delete_empty,
3341         .walk           = fl_walk,
3342         .reoffload      = fl_reoffload,
3343         .hw_add         = fl_hw_add,
3344         .hw_del         = fl_hw_del,
3345         .dump           = fl_dump,
3346         .terse_dump     = fl_terse_dump,
3347         .bind_class     = fl_bind_class,
3348         .tmplt_create   = fl_tmplt_create,
3349         .tmplt_destroy  = fl_tmplt_destroy,
3350         .tmplt_dump     = fl_tmplt_dump,
3351         .owner          = THIS_MODULE,
3352         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3353 };
3354
3355 static int __init cls_fl_init(void)
3356 {
3357         return register_tcf_proto_ops(&cls_fl_ops);
3358 }
3359
3360 static void __exit cls_fl_exit(void)
3361 {
3362         unregister_tcf_proto_ops(&cls_fl_ops);
3363 }
3364
3365 module_init(cls_fl_init);
3366 module_exit(cls_fl_exit);
3367
3368 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3369 MODULE_DESCRIPTION("Flower classifier");
3370 MODULE_LICENSE("GPL v2");