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 <[email protected]>
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/rhashtable.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/refcount.h>
  14
  15 #include <linux/if_ether.h>
  16 #include <linux/in6.h>
  17 #include <linux/ip.h>
  18 #include <linux/mpls.h>
  19
  20 #include <net/sch_generic.h>
  21 #include <net/pkt_cls.h>
  22 #include <net/ip.h>
  23 #include <net/flow_dissector.h>
  24 #include <net/geneve.h>
  25 #include <net/vxlan.h>
  26 #include <net/erspan.h>
  27
  28 #include <net/dst.h>
  29 #include <net/dst_metadata.h>
  30
  31 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  32
  33 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  34                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  35 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  36                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  37
  38 struct fl_flow_key {
  39         struct flow_dissector_key_meta meta;
  40         struct flow_dissector_key_control control;
  41         struct flow_dissector_key_control enc_control;
  42         struct flow_dissector_key_basic basic;
  43         struct flow_dissector_key_eth_addrs eth;
  44         struct flow_dissector_key_vlan vlan;
  45         struct flow_dissector_key_vlan cvlan;
  46         union {
  47                 struct flow_dissector_key_ipv4_addrs ipv4;
  48                 struct flow_dissector_key_ipv6_addrs ipv6;
  49         };
  50         struct flow_dissector_key_ports tp;
  51         struct flow_dissector_key_icmp icmp;
  52         struct flow_dissector_key_arp arp;
  53         struct flow_dissector_key_keyid enc_key_id;
  54         union {
  55                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  56                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  57         };
  58         struct flow_dissector_key_ports enc_tp;
  59         struct flow_dissector_key_mpls mpls;
  60         struct flow_dissector_key_tcp tcp;
  61         struct flow_dissector_key_ip ip;
  62         struct flow_dissector_key_ip enc_ip;
  63         struct flow_dissector_key_enc_opts enc_opts;
  64         union {
  65                 struct flow_dissector_key_ports tp;
  66                 struct {
  67                         struct flow_dissector_key_ports tp_min;
  68                         struct flow_dissector_key_ports tp_max;
  69                 };
  70         } tp_range;
  71         struct flow_dissector_key_ct ct;
  72         struct flow_dissector_key_hash hash;
  73 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  74
  75 struct fl_flow_mask_range {
  76         unsigned short int start;
  77         unsigned short int end;
  78 };
  79
  80 struct fl_flow_mask {
  81         struct fl_flow_key key;
  82         struct fl_flow_mask_range range;
  83         u32 flags;
  84         struct rhash_head ht_node;
  85         struct rhashtable ht;
  86         struct rhashtable_params filter_ht_params;
  87         struct flow_dissector dissector;
  88         struct list_head filters;
  89         struct rcu_work rwork;
  90         struct list_head list;
  91         refcount_t refcnt;
  92 };
  93
  94 struct fl_flow_tmplt {
  95         struct fl_flow_key dummy_key;
  96         struct fl_flow_key mask;
  97         struct flow_dissector dissector;
  98         struct tcf_chain *chain;
  99 };
 100
 101 struct cls_fl_head {
 102         struct rhashtable ht;
 103         spinlock_t masks_lock; /* Protect masks list */
 104         struct list_head masks;
 105         struct list_head hw_filters;
 106         struct rcu_work rwork;
 107         struct idr handle_idr;
 108 };
 109
 110 struct cls_fl_filter {
 111         struct fl_flow_mask *mask;
 112         struct rhash_head ht_node;
 113         struct fl_flow_key mkey;
 114         struct tcf_exts exts;
 115         struct tcf_result res;
 116         struct fl_flow_key key;
 117         struct list_head list;
 118         struct list_head hw_list;
 119         u32 handle;
 120         u32 flags;
 121         u32 in_hw_count;
 122         struct rcu_work rwork;
 123         struct net_device *hw_dev;
 124         /* Flower classifier is unlocked, which means that its reference counter
 125          * can be changed concurrently without any kind of external
 126          * synchronization. Use atomic reference counter to be concurrency-safe.
 127          */
 128         refcount_t refcnt;
 129         bool deleted;
 130 };
 131
 132 static const struct rhashtable_params mask_ht_params = {
 133         .key_offset = offsetof(struct fl_flow_mask, key),
 134         .key_len = sizeof(struct fl_flow_key),
 135         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 136         .automatic_shrinking = true,
 137 };
 138
 139 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 140 {
 141         return mask->range.end - mask->range.start;
 142 }
 143
 144 static void fl_mask_update_range(struct fl_flow_mask *mask)
 145 {
 146         const u8 *bytes = (const u8 *) &mask->key;
 147         size_t size = sizeof(mask->key);
 148         size_t i, first = 0, last;
 149
 150         for (i = 0; i < size; i++) {
 151                 if (bytes[i]) {
 152                         first = i;
 153                         break;
 154                 }
 155         }
 156         last = first;
 157         for (i = size - 1; i != first; i--) {
 158                 if (bytes[i]) {
 159                         last = i;
 160                         break;
 161                 }
 162         }
 163         mask->range.start = rounddown(first, sizeof(long));
 164         mask->range.end = roundup(last + 1, sizeof(long));
 165 }
 166
 167 static void *fl_key_get_start(struct fl_flow_key *key,
 168                               const struct fl_flow_mask *mask)
 169 {
 170         return (u8 *) key + mask->range.start;
 171 }
 172
 173 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 174                               struct fl_flow_mask *mask)
 175 {
 176         const long *lkey = fl_key_get_start(key, mask);
 177         const long *lmask = fl_key_get_start(&mask->key, mask);
 178         long *lmkey = fl_key_get_start(mkey, mask);
 179         int i;
 180
 181         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 182                 *lmkey++ = *lkey++ & *lmask++;
 183 }
 184
 185 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 186                                struct fl_flow_mask *mask)
 187 {
 188         const long *lmask = fl_key_get_start(&mask->key, mask);
 189         const long *ltmplt;
 190         int i;
 191
 192         if (!tmplt)
 193                 return true;
 194         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 195         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 196                 if (~*ltmplt++ & *lmask++)
 197                         return false;
 198         }
 199         return true;
 200 }
 201
 202 static void fl_clear_masked_range(struct fl_flow_key *key,
 203                                   struct fl_flow_mask *mask)
 204 {
 205         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 206 }
 207
 208 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 209                                   struct fl_flow_key *key,
 210                                   struct fl_flow_key *mkey)
 211 {
 212         u16 min_mask, max_mask, min_val, max_val;
 213
 214         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 215         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 216         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 217         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 218
 219         if (min_mask && max_mask) {
 220                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 221                     ntohs(key->tp_range.tp.dst) > max_val)
 222                         return false;
 223
 224                 /* skb does not have min and max values */
 225                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 226                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 227         }
 228         return true;
 229 }
 230
 231 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 232                                   struct fl_flow_key *key,
 233                                   struct fl_flow_key *mkey)
 234 {
 235         u16 min_mask, max_mask, min_val, max_val;
 236
 237         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 238         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 239         min_val = ntohs(filter->key.tp_range.tp_min.src);
 240         max_val = ntohs(filter->key.tp_range.tp_max.src);
 241
 242         if (min_mask && max_mask) {
 243                 if (ntohs(key->tp_range.tp.src) < min_val ||
 244                     ntohs(key->tp_range.tp.src) > max_val)
 245                         return false;
 246
 247                 /* skb does not have min and max values */
 248                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 249                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 250         }
 251         return true;
 252 }
 253
 254 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 255                                          struct fl_flow_key *mkey)
 256 {
 257         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 258                                       mask->filter_ht_params);
 259 }
 260
 261 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 262                                              struct fl_flow_key *mkey,
 263                                              struct fl_flow_key *key)
 264 {
 265         struct cls_fl_filter *filter, *f;
 266
 267         list_for_each_entry_rcu(filter, &mask->filters, list) {
 268                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 269                         continue;
 270
 271                 if (!fl_range_port_src_cmp(filter, key, mkey))
 272                         continue;
 273
 274                 f = __fl_lookup(mask, mkey);
 275                 if (f)
 276                         return f;
 277         }
 278         return NULL;
 279 }
 280
 281 static noinline_for_stack
 282 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 283 {
 284         struct fl_flow_key mkey;
 285
 286         fl_set_masked_key(&mkey, key, mask);
 287         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 288                 return fl_lookup_range(mask, &mkey, key);
 289
 290         return __fl_lookup(mask, &mkey);
 291 }
 292
 293 static u16 fl_ct_info_to_flower_map[] = {
 294         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 295                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 296         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 297                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 298         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 299                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED |
 300                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 301         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 302                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED |
 303                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 304         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 305                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 306 };
 307
 308 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 309                        struct tcf_result *res)
 310 {
 311         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 312         bool post_ct = qdisc_skb_cb(skb)->post_ct;
 313         struct fl_flow_key skb_key;
 314         struct fl_flow_mask *mask;
 315         struct cls_fl_filter *f;
 316
 317         list_for_each_entry_rcu(mask, &head->masks, list) {
 318                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 319                 fl_clear_masked_range(&skb_key, mask);
 320
 321                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 322                 /* skb_flow_dissect() does not set n_proto in case an unknown
 323                  * protocol, so do it rather here.
 324                  */
 325                 skb_key.basic.n_proto = skb_protocol(skb, false);
 326                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 327                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 328                                     fl_ct_info_to_flower_map,
 329                                     ARRAY_SIZE(fl_ct_info_to_flower_map),
 330                                     post_ct);
 331                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 332                 skb_flow_dissect(skb, &mask->dissector, &skb_key,
 333                                  FLOW_DISSECTOR_F_STOP_BEFORE_ENCAP);
 334
 335                 f = fl_mask_lookup(mask, &skb_key);
 336                 if (f && !tc_skip_sw(f->flags)) {
 337                         *res = f->res;
 338                         return tcf_exts_exec(skb, &f->exts, res);
 339                 }
 340         }
 341         return -1;
 342 }
 343
 344 static int fl_init(struct tcf_proto *tp)
 345 {
 346         struct cls_fl_head *head;
 347
 348         head = kzalloc(sizeof(*head), GFP_KERNEL);
 349         if (!head)
 350                 return -ENOBUFS;
 351
 352         spin_lock_init(&head->masks_lock);
 353         INIT_LIST_HEAD_RCU(&head->masks);
 354         INIT_LIST_HEAD(&head->hw_filters);
 355         rcu_assign_pointer(tp->root, head);
 356         idr_init(&head->handle_idr);
 357
 358         return rhashtable_init(&head->ht, &mask_ht_params);
 359 }
 360
 361 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 362 {
 363         /* temporary masks don't have their filters list and ht initialized */
 364         if (mask_init_done) {
 365                 WARN_ON(!list_empty(&mask->filters));
 366                 rhashtable_destroy(&mask->ht);
 367         }
 368         kfree(mask);
 369 }
 370
 371 static void fl_mask_free_work(struct work_struct *work)
 372 {
 373         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 374                                                  struct fl_flow_mask, rwork);
 375
 376         fl_mask_free(mask, true);
 377 }
 378
 379 static void fl_uninit_mask_free_work(struct work_struct *work)
 380 {
 381         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 382                                                  struct fl_flow_mask, rwork);
 383
 384         fl_mask_free(mask, false);
 385 }
 386
 387 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 388 {
 389         if (!refcount_dec_and_test(&mask->refcnt))
 390                 return false;
 391
 392         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 393
 394         spin_lock(&head->masks_lock);
 395         list_del_rcu(&mask->list);
 396         spin_unlock(&head->masks_lock);
 397
 398         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 399
 400         return true;
 401 }
 402
 403 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 404 {
 405         /* Flower classifier only changes root pointer during init and destroy.
 406          * Users must obtain reference to tcf_proto instance before calling its
 407          * API, so tp->root pointer is protected from concurrent call to
 408          * fl_destroy() by reference counting.
 409          */
 410         return rcu_dereference_raw(tp->root);
 411 }
 412
 413 static void __fl_destroy_filter(struct cls_fl_filter *f)
 414 {
 415         tcf_exts_destroy(&f->exts);
 416         tcf_exts_put_net(&f->exts);
 417         kfree(f);
 418 }
 419
 420 static void fl_destroy_filter_work(struct work_struct *work)
 421 {
 422         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 423                                         struct cls_fl_filter, rwork);
 424
 425         __fl_destroy_filter(f);
 426 }
 427
 428 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 429                                  bool rtnl_held, struct netlink_ext_ack *extack)
 430 {
 431         struct tcf_block *block = tp->chain->block;
 432         struct flow_cls_offload cls_flower = {};
 433
 434         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 435         cls_flower.command = FLOW_CLS_DESTROY;
 436         cls_flower.cookie = (unsigned long) f;
 437
 438         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 439                             &f->flags, &f->in_hw_count, rtnl_held);
 440
 441 }
 442
 443 static int fl_hw_replace_filter(struct tcf_proto *tp,
 444                                 struct cls_fl_filter *f, bool rtnl_held,
 445                                 struct netlink_ext_ack *extack)
 446 {
 447         struct tcf_block *block = tp->chain->block;
 448         struct flow_cls_offload cls_flower = {};
 449         bool skip_sw = tc_skip_sw(f->flags);
 450         int err = 0;
 451
 452         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 453         if (!cls_flower.rule)
 454                 return -ENOMEM;
 455
 456         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 457         cls_flower.command = FLOW_CLS_REPLACE;
 458         cls_flower.cookie = (unsigned long) f;
 459         cls_flower.rule->match.dissector = &f->mask->dissector;
 460         cls_flower.rule->match.mask = &f->mask->key;
 461         cls_flower.rule->match.key = &f->mkey;
 462         cls_flower.classid = f->res.classid;
 463
 464         err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
 465         if (err) {
 466                 kfree(cls_flower.rule);
 467                 if (skip_sw) {
 468                         NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
 469                         return err;
 470                 }
 471                 return 0;
 472         }
 473
 474         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 475                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 476         tc_cleanup_flow_action(&cls_flower.rule->action);
 477         kfree(cls_flower.rule);
 478
 479         if (err) {
 480                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 481                 return err;
 482         }
 483
 484         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 485                 return -EINVAL;
 486
 487         return 0;
 488 }
 489
 490 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 491                                bool rtnl_held)
 492 {
 493         struct tcf_block *block = tp->chain->block;
 494         struct flow_cls_offload cls_flower = {};
 495
 496         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 497         cls_flower.command = FLOW_CLS_STATS;
 498         cls_flower.cookie = (unsigned long) f;
 499         cls_flower.classid = f->res.classid;
 500
 501         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 502                          rtnl_held);
 503
 504         tcf_exts_stats_update(&f->exts, cls_flower.stats.bytes,
 505                               cls_flower.stats.pkts,
 506                               cls_flower.stats.drops,
 507                               cls_flower.stats.lastused,
 508                               cls_flower.stats.used_hw_stats,
 509                               cls_flower.stats.used_hw_stats_valid);
 510 }
 511
 512 static void __fl_put(struct cls_fl_filter *f)
 513 {
 514         if (!refcount_dec_and_test(&f->refcnt))
 515                 return;
 516
 517         if (tcf_exts_get_net(&f->exts))
 518                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 519         else
 520                 __fl_destroy_filter(f);
 521 }
 522
 523 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 524 {
 525         struct cls_fl_filter *f;
 526
 527         rcu_read_lock();
 528         f = idr_find(&head->handle_idr, handle);
 529         if (f && !refcount_inc_not_zero(&f->refcnt))
 530                 f = NULL;
 531         rcu_read_unlock();
 532
 533         return f;
 534 }
 535
 536 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 537                        bool *last, bool rtnl_held,
 538                        struct netlink_ext_ack *extack)
 539 {
 540         struct cls_fl_head *head = fl_head_dereference(tp);
 541
 542         *last = false;
 543
 544         spin_lock(&tp->lock);
 545         if (f->deleted) {
 546                 spin_unlock(&tp->lock);
 547                 return -ENOENT;
 548         }
 549
 550         f->deleted = true;
 551         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 552                                f->mask->filter_ht_params);
 553         idr_remove(&head->handle_idr, f->handle);
 554         list_del_rcu(&f->list);
 555         spin_unlock(&tp->lock);
 556
 557         *last = fl_mask_put(head, f->mask);
 558         if (!tc_skip_hw(f->flags))
 559                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 560         tcf_unbind_filter(tp, &f->res);
 561         __fl_put(f);
 562
 563         return 0;
 564 }
 565
 566 static void fl_destroy_sleepable(struct work_struct *work)
 567 {
 568         struct cls_fl_head *head = container_of(to_rcu_work(work),
 569                                                 struct cls_fl_head,
 570                                                 rwork);
 571
 572         rhashtable_destroy(&head->ht);
 573         kfree(head);
 574         module_put(THIS_MODULE);
 575 }
 576
 577 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 578                        struct netlink_ext_ack *extack)
 579 {
 580         struct cls_fl_head *head = fl_head_dereference(tp);
 581         struct fl_flow_mask *mask, *next_mask;
 582         struct cls_fl_filter *f, *next;
 583         bool last;
 584
 585         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 586                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 587                         __fl_delete(tp, f, &last, rtnl_held, extack);
 588                         if (last)
 589                                 break;
 590                 }
 591         }
 592         idr_destroy(&head->handle_idr);
 593
 594         __module_get(THIS_MODULE);
 595         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 596 }
 597
 598 static void fl_put(struct tcf_proto *tp, void *arg)
 599 {
 600         struct cls_fl_filter *f = arg;
 601
 602         __fl_put(f);
 603 }
 604
 605 static void *fl_get(struct tcf_proto *tp, u32 handle)
 606 {
 607         struct cls_fl_head *head = fl_head_dereference(tp);
 608
 609         return __fl_get(head, handle);
 610 }
 611
 612 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 613         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 614         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 615         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 616                                             .len = IFNAMSIZ },
 617         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 618         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 619         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 620         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 621         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 622         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 623         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 624         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 625         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 626         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 627         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 628         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 629         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 630         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 631         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 632         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 633         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 634         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 635         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 636         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 637         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 638         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 639         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 640         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 641         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 642         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 643         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 644         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 645         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 646         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 647         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 648         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 649         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 650         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 660         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 661         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 662         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 663         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 664         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 665         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 666         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 670         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 671         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 672         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 673         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 674         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 675         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 676         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 677         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 678         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 679         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 680         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 681         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 682         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 683         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 684         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 685         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 686         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 687         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 688         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 690         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 691         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 692         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 693         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 694         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 695         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 696         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 697         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 698         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 699         [TCA_FLOWER_KEY_CT_STATE]       =
 700                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 701         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 702                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 703         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 704         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 705         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 706         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 707         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 708                                             .len = 128 / BITS_PER_BYTE },
 709         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 710                                             .len = 128 / BITS_PER_BYTE },
 711         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 712         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 713         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 714
 715 };
 716
 717 static const struct nla_policy
 718 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 719         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 720                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 721         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 722         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 723         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 724 };
 725
 726 static const struct nla_policy
 727 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 728         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 729         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 730         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 731                                                        .len = 128 },
 732 };
 733
 734 static const struct nla_policy
 735 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 736         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 737 };
 738
 739 static const struct nla_policy
 740 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 741         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 742         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 743         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 744         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 745 };
 746
 747 static const struct nla_policy
 748 mpls_stack_entry_policy[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1] = {
 749         [TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]    = { .type = NLA_U8 },
 750         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]      = { .type = NLA_U8 },
 751         [TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]      = { .type = NLA_U8 },
 752         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]       = { .type = NLA_U8 },
 753         [TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]    = { .type = NLA_U32 },
 754 };
 755
 756 static void fl_set_key_val(struct nlattr **tb,
 757                            void *val, int val_type,
 758                            void *mask, int mask_type, int len)
 759 {
 760         if (!tb[val_type])
 761                 return;
 762         nla_memcpy(val, tb[val_type], len);
 763         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 764                 memset(mask, 0xff, len);
 765         else
 766                 nla_memcpy(mask, tb[mask_type], len);
 767 }
 768
 769 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 770                                  struct fl_flow_key *mask,
 771                                  struct netlink_ext_ack *extack)
 772 {
 773         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 774                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 775                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 776         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 777                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 778                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 779         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 780                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 781                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 782         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 783                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 784                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 785
 786         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 787             ntohs(key->tp_range.tp_max.dst) <=
 788             ntohs(key->tp_range.tp_min.dst)) {
 789                 NL_SET_ERR_MSG_ATTR(extack,
 790                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 791                                     "Invalid destination port range (min must be strictly smaller than max)");
 792                 return -EINVAL;
 793         }
 794         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 795             ntohs(key->tp_range.tp_max.src) <=
 796             ntohs(key->tp_range.tp_min.src)) {
 797                 NL_SET_ERR_MSG_ATTR(extack,
 798                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 799                                     "Invalid source port range (min must be strictly smaller than max)");
 800                 return -EINVAL;
 801         }
 802
 803         return 0;
 804 }
 805
 806 static int fl_set_key_mpls_lse(const struct nlattr *nla_lse,
 807                                struct flow_dissector_key_mpls *key_val,
 808                                struct flow_dissector_key_mpls *key_mask,
 809                                struct netlink_ext_ack *extack)
 810 {
 811         struct nlattr *tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1];
 812         struct flow_dissector_mpls_lse *lse_mask;
 813         struct flow_dissector_mpls_lse *lse_val;
 814         u8 lse_index;
 815         u8 depth;
 816         int err;
 817
 818         err = nla_parse_nested(tb, TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX, nla_lse,
 819                                mpls_stack_entry_policy, extack);
 820         if (err < 0)
 821                 return err;
 822
 823         if (!tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]) {
 824                 NL_SET_ERR_MSG(extack, "Missing MPLS option \"depth\"");
 825                 return -EINVAL;
 826         }
 827
 828         depth = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]);
 829
 830         /* LSE depth starts at 1, for consistency with terminology used by
 831          * RFC 3031 (section 3.9), where depth 0 refers to unlabeled packets.
 832          */
 833         if (depth < 1 || depth > FLOW_DIS_MPLS_MAX) {
 834                 NL_SET_ERR_MSG_ATTR(extack,
 835                                     tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH],
 836                                     "Invalid MPLS depth");
 837                 return -EINVAL;
 838         }
 839         lse_index = depth - 1;
 840
 841         dissector_set_mpls_lse(key_val, lse_index);
 842         dissector_set_mpls_lse(key_mask, lse_index);
 843
 844         lse_val = &key_val->ls[lse_index];
 845         lse_mask = &key_mask->ls[lse_index];
 846
 847         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]) {
 848                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]);
 849                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 850         }
 851         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]) {
 852                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]);
 853
 854                 if (bos & ~MPLS_BOS_MASK) {
 855                         NL_SET_ERR_MSG_ATTR(extack,
 856                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS],
 857                                             "Bottom Of Stack (BOS) must be 0 or 1");
 858                         return -EINVAL;
 859                 }
 860                 lse_val->mpls_bos = bos;
 861                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 862         }
 863         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]) {
 864                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]);
 865
 866                 if (tc & ~MPLS_TC_MASK) {
 867                         NL_SET_ERR_MSG_ATTR(extack,
 868                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC],
 869                                             "Traffic Class (TC) must be between 0 and 7");
 870                         return -EINVAL;
 871                 }
 872                 lse_val->mpls_tc = tc;
 873                 lse_mask->mpls_tc = MPLS_TC_MASK;
 874         }
 875         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]) {
 876                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]);
 877
 878                 if (label & ~MPLS_LABEL_MASK) {
 879                         NL_SET_ERR_MSG_ATTR(extack,
 880                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL],
 881                                             "Label must be between 0 and 1048575");
 882                         return -EINVAL;
 883                 }
 884                 lse_val->mpls_label = label;
 885                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 886         }
 887
 888         return 0;
 889 }
 890
 891 static int fl_set_key_mpls_opts(const struct nlattr *nla_mpls_opts,
 892                                 struct flow_dissector_key_mpls *key_val,
 893                                 struct flow_dissector_key_mpls *key_mask,
 894                                 struct netlink_ext_ack *extack)
 895 {
 896         struct nlattr *nla_lse;
 897         int rem;
 898         int err;
 899
 900         if (!(nla_mpls_opts->nla_type & NLA_F_NESTED)) {
 901                 NL_SET_ERR_MSG_ATTR(extack, nla_mpls_opts,
 902                                     "NLA_F_NESTED is missing");
 903                 return -EINVAL;
 904         }
 905
 906         nla_for_each_nested(nla_lse, nla_mpls_opts, rem) {
 907                 if (nla_type(nla_lse) != TCA_FLOWER_KEY_MPLS_OPTS_LSE) {
 908                         NL_SET_ERR_MSG_ATTR(extack, nla_lse,
 909                                             "Invalid MPLS option type");
 910                         return -EINVAL;
 911                 }
 912
 913                 err = fl_set_key_mpls_lse(nla_lse, key_val, key_mask, extack);
 914                 if (err < 0)
 915                         return err;
 916         }
 917         if (rem) {
 918                 NL_SET_ERR_MSG(extack,
 919                                "Bytes leftover after parsing MPLS options");
 920                 return -EINVAL;
 921         }
 922
 923         return 0;
 924 }
 925
 926 static int fl_set_key_mpls(struct nlattr **tb,
 927                            struct flow_dissector_key_mpls *key_val,
 928                            struct flow_dissector_key_mpls *key_mask,
 929                            struct netlink_ext_ack *extack)
 930 {
 931         struct flow_dissector_mpls_lse *lse_mask;
 932         struct flow_dissector_mpls_lse *lse_val;
 933
 934         if (tb[TCA_FLOWER_KEY_MPLS_OPTS]) {
 935                 if (tb[TCA_FLOWER_KEY_MPLS_TTL] ||
 936                     tb[TCA_FLOWER_KEY_MPLS_BOS] ||
 937                     tb[TCA_FLOWER_KEY_MPLS_TC] ||
 938                     tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 939                         NL_SET_ERR_MSG_ATTR(extack,
 940                                             tb[TCA_FLOWER_KEY_MPLS_OPTS],
 941                                             "MPLS label, Traffic Class, Bottom Of Stack and Time To Live must be encapsulated in the MPLS options attribute");
 942                         return -EBADMSG;
 943                 }
 944
 945                 return fl_set_key_mpls_opts(tb[TCA_FLOWER_KEY_MPLS_OPTS],
 946                                             key_val, key_mask, extack);
 947         }
 948
 949         lse_val = &key_val->ls[0];
 950         lse_mask = &key_mask->ls[0];
 951
 952         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 953                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 954                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 955                 dissector_set_mpls_lse(key_val, 0);
 956                 dissector_set_mpls_lse(key_mask, 0);
 957         }
 958         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 959                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 960
 961                 if (bos & ~MPLS_BOS_MASK) {
 962                         NL_SET_ERR_MSG_ATTR(extack,
 963                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
 964                                             "Bottom Of Stack (BOS) must be 0 or 1");
 965                         return -EINVAL;
 966                 }
 967                 lse_val->mpls_bos = bos;
 968                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 969                 dissector_set_mpls_lse(key_val, 0);
 970                 dissector_set_mpls_lse(key_mask, 0);
 971         }
 972         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 973                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 974
 975                 if (tc & ~MPLS_TC_MASK) {
 976                         NL_SET_ERR_MSG_ATTR(extack,
 977                                             tb[TCA_FLOWER_KEY_MPLS_TC],
 978                                             "Traffic Class (TC) must be between 0 and 7");
 979                         return -EINVAL;
 980                 }
 981                 lse_val->mpls_tc = tc;
 982                 lse_mask->mpls_tc = MPLS_TC_MASK;
 983                 dissector_set_mpls_lse(key_val, 0);
 984                 dissector_set_mpls_lse(key_mask, 0);
 985         }
 986         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 987                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 988
 989                 if (label & ~MPLS_LABEL_MASK) {
 990                         NL_SET_ERR_MSG_ATTR(extack,
 991                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
 992                                             "Label must be between 0 and 1048575");
 993                         return -EINVAL;
 994                 }
 995                 lse_val->mpls_label = label;
 996                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 997                 dissector_set_mpls_lse(key_val, 0);
 998                 dissector_set_mpls_lse(key_mask, 0);
 999         }
1000         return 0;
1001 }
1002
1003 static void fl_set_key_vlan(struct nlattr **tb,
1004                             __be16 ethertype,
1005                             int vlan_id_key, int vlan_prio_key,
1006                             struct flow_dissector_key_vlan *key_val,
1007                             struct flow_dissector_key_vlan *key_mask)
1008 {
1009 #define VLAN_PRIORITY_MASK      0x7
1010
1011         if (tb[vlan_id_key]) {
1012                 key_val->vlan_id =
1013                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1014                 key_mask->vlan_id = VLAN_VID_MASK;
1015         }
1016         if (tb[vlan_prio_key]) {
1017                 key_val->vlan_priority =
1018                         nla_get_u8(tb[vlan_prio_key]) &
1019                         VLAN_PRIORITY_MASK;
1020                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1021         }
1022         key_val->vlan_tpid = ethertype;
1023         key_mask->vlan_tpid = cpu_to_be16(~0);
1024 }
1025
1026 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1027                             u32 *dissector_key, u32 *dissector_mask,
1028                             u32 flower_flag_bit, u32 dissector_flag_bit)
1029 {
1030         if (flower_mask & flower_flag_bit) {
1031                 *dissector_mask |= dissector_flag_bit;
1032                 if (flower_key & flower_flag_bit)
1033                         *dissector_key |= dissector_flag_bit;
1034         }
1035 }
1036
1037 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1038                             u32 *flags_mask, struct netlink_ext_ack *extack)
1039 {
1040         u32 key, mask;
1041
1042         /* mask is mandatory for flags */
1043         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1044                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1045                 return -EINVAL;
1046         }
1047
1048         key = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS]));
1049         mask = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1050
1051         *flags_key  = 0;
1052         *flags_mask = 0;
1053
1054         fl_set_key_flag(key, mask, flags_key, flags_mask,
1055                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1056         fl_set_key_flag(key, mask, flags_key, flags_mask,
1057                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1058                         FLOW_DIS_FIRST_FRAG);
1059
1060         return 0;
1061 }
1062
1063 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1064                           struct flow_dissector_key_ip *key,
1065                           struct flow_dissector_key_ip *mask)
1066 {
1067         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1068         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1069         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1070         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1071
1072         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1073         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1074 }
1075
1076 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1077                              int depth, int option_len,
1078                              struct netlink_ext_ack *extack)
1079 {
1080         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1081         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1082         struct geneve_opt *opt;
1083         int err, data_len = 0;
1084
1085         if (option_len > sizeof(struct geneve_opt))
1086                 data_len = option_len - sizeof(struct geneve_opt);
1087
1088         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1089         memset(opt, 0xff, option_len);
1090         opt->length = data_len / 4;
1091         opt->r1 = 0;
1092         opt->r2 = 0;
1093         opt->r3 = 0;
1094
1095         /* If no mask has been prodived we assume an exact match. */
1096         if (!depth)
1097                 return sizeof(struct geneve_opt) + data_len;
1098
1099         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1100                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1101                 return -EINVAL;
1102         }
1103
1104         err = nla_parse_nested_deprecated(tb,
1105                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1106                                           nla, geneve_opt_policy, extack);
1107         if (err < 0)
1108                 return err;
1109
1110         /* We are not allowed to omit any of CLASS, TYPE or DATA
1111          * fields from the key.
1112          */
1113         if (!option_len &&
1114             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1115              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1116              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1117                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1118                 return -EINVAL;
1119         }
1120
1121         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1122          * for the mask.
1123          */
1124         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1125                 int new_len = key->enc_opts.len;
1126
1127                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1128                 data_len = nla_len(data);
1129                 if (data_len < 4) {
1130                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1131                         return -ERANGE;
1132                 }
1133                 if (data_len % 4) {
1134                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1135                         return -ERANGE;
1136                 }
1137
1138                 new_len += sizeof(struct geneve_opt) + data_len;
1139                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1140                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1141                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1142                         return -ERANGE;
1143                 }
1144                 opt->length = data_len / 4;
1145                 memcpy(opt->opt_data, nla_data(data), data_len);
1146         }
1147
1148         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1149                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1150                 opt->opt_class = nla_get_be16(class);
1151         }
1152
1153         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1154                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1155                 opt->type = nla_get_u8(type);
1156         }
1157
1158         return sizeof(struct geneve_opt) + data_len;
1159 }
1160
1161 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1162                             int depth, int option_len,
1163                             struct netlink_ext_ack *extack)
1164 {
1165         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1166         struct vxlan_metadata *md;
1167         int err;
1168
1169         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1170         memset(md, 0xff, sizeof(*md));
1171
1172         if (!depth)
1173                 return sizeof(*md);
1174
1175         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1176                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1177                 return -EINVAL;
1178         }
1179
1180         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1181                                vxlan_opt_policy, extack);
1182         if (err < 0)
1183                 return err;
1184
1185         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1186                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1187                 return -EINVAL;
1188         }
1189
1190         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1191                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1192                 md->gbp &= VXLAN_GBP_MASK;
1193         }
1194
1195         return sizeof(*md);
1196 }
1197
1198 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1199                              int depth, int option_len,
1200                              struct netlink_ext_ack *extack)
1201 {
1202         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1203         struct erspan_metadata *md;
1204         int err;
1205
1206         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1207         memset(md, 0xff, sizeof(*md));
1208         md->version = 1;
1209
1210         if (!depth)
1211                 return sizeof(*md);
1212
1213         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1214                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1215                 return -EINVAL;
1216         }
1217
1218         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1219                                erspan_opt_policy, extack);
1220         if (err < 0)
1221                 return err;
1222
1223         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1224                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1225                 return -EINVAL;
1226         }
1227
1228         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1229                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1230
1231         if (md->version == 1) {
1232                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1233                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1234                         return -EINVAL;
1235                 }
1236                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1237                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1238                         memset(&md->u, 0x00, sizeof(md->u));
1239                         md->u.index = nla_get_be32(nla);
1240                 }
1241         } else if (md->version == 2) {
1242                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1243                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1244                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1245                         return -EINVAL;
1246                 }
1247                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1248                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1249                         md->u.md2.dir = nla_get_u8(nla);
1250                 }
1251                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1252                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1253                         set_hwid(&md->u.md2, nla_get_u8(nla));
1254                 }
1255         } else {
1256                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1257                 return -EINVAL;
1258         }
1259
1260         return sizeof(*md);
1261 }
1262
1263 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1264                           struct fl_flow_key *mask,
1265                           struct netlink_ext_ack *extack)
1266 {
1267         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1268         int err, option_len, key_depth, msk_depth = 0;
1269
1270         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1271                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1272                                              enc_opts_policy, extack);
1273         if (err)
1274                 return err;
1275
1276         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1277
1278         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1279                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1280                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1281                                                      enc_opts_policy, extack);
1282                 if (err)
1283                         return err;
1284
1285                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1286                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1287                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1288                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1289                         return -EINVAL;
1290                 }
1291         }
1292
1293         nla_for_each_attr(nla_opt_key, nla_enc_key,
1294                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1295                 switch (nla_type(nla_opt_key)) {
1296                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1297                         if (key->enc_opts.dst_opt_type &&
1298                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1299                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1300                                 return -EINVAL;
1301                         }
1302                         option_len = 0;
1303                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1304                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1305                                                        key_depth, option_len,
1306                                                        extack);
1307                         if (option_len < 0)
1308                                 return option_len;
1309
1310                         key->enc_opts.len += option_len;
1311                         /* At the same time we need to parse through the mask
1312                          * in order to verify exact and mask attribute lengths.
1313                          */
1314                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1315                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1316                                                        msk_depth, option_len,
1317                                                        extack);
1318                         if (option_len < 0)
1319                                 return option_len;
1320
1321                         mask->enc_opts.len += option_len;
1322                         if (key->enc_opts.len != mask->enc_opts.len) {
1323                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1324                                 return -EINVAL;
1325                         }
1326                         break;
1327                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1328                         if (key->enc_opts.dst_opt_type) {
1329                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1330                                 return -EINVAL;
1331                         }
1332                         option_len = 0;
1333                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1334                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1335                                                       key_depth, option_len,
1336                                                       extack);
1337                         if (option_len < 0)
1338                                 return option_len;
1339
1340                         key->enc_opts.len += option_len;
1341                         /* At the same time we need to parse through the mask
1342                          * in order to verify exact and mask attribute lengths.
1343                          */
1344                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1345                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1346                                                       msk_depth, option_len,
1347                                                       extack);
1348                         if (option_len < 0)
1349                                 return option_len;
1350
1351                         mask->enc_opts.len += option_len;
1352                         if (key->enc_opts.len != mask->enc_opts.len) {
1353                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1354                                 return -EINVAL;
1355                         }
1356                         break;
1357                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1358                         if (key->enc_opts.dst_opt_type) {
1359                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1360                                 return -EINVAL;
1361                         }
1362                         option_len = 0;
1363                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1364                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1365                                                        key_depth, option_len,
1366                                                        extack);
1367                         if (option_len < 0)
1368                                 return option_len;
1369
1370                         key->enc_opts.len += option_len;
1371                         /* At the same time we need to parse through the mask
1372                          * in order to verify exact and mask attribute lengths.
1373                          */
1374                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1375                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1376                                                        msk_depth, option_len,
1377                                                        extack);
1378                         if (option_len < 0)
1379                                 return option_len;
1380
1381                         mask->enc_opts.len += option_len;
1382                         if (key->enc_opts.len != mask->enc_opts.len) {
1383                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1384                                 return -EINVAL;
1385                         }
1386                         break;
1387                 default:
1388                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1389                         return -EINVAL;
1390                 }
1391
1392                 if (!msk_depth)
1393                         continue;
1394
1395                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1396                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1397                         return -EINVAL;
1398                 }
1399                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1400         }
1401
1402         return 0;
1403 }
1404
1405 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1406                                 struct netlink_ext_ack *extack)
1407 {
1408         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1409                 NL_SET_ERR_MSG_ATTR(extack, tb,
1410                                     "no trk, so no other flag can be set");
1411                 return -EINVAL;
1412         }
1413
1414         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1415             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1416                 NL_SET_ERR_MSG_ATTR(extack, tb,
1417                                     "new and est are mutually exclusive");
1418                 return -EINVAL;
1419         }
1420
1421         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1422             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1423                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1424                 NL_SET_ERR_MSG_ATTR(extack, tb,
1425                                     "when inv is set, only trk may be set");
1426                 return -EINVAL;
1427         }
1428
1429         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1430             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1431                 NL_SET_ERR_MSG_ATTR(extack, tb,
1432                                     "new and rpl are mutually exclusive");
1433                 return -EINVAL;
1434         }
1435
1436         return 0;
1437 }
1438
1439 static int fl_set_key_ct(struct nlattr **tb,
1440                          struct flow_dissector_key_ct *key,
1441                          struct flow_dissector_key_ct *mask,
1442                          struct netlink_ext_ack *extack)
1443 {
1444         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1445                 int err;
1446
1447                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1448                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1449                         return -EOPNOTSUPP;
1450                 }
1451                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1452                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1453                                sizeof(key->ct_state));
1454
1455                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1456                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1457                                            extack);
1458                 if (err)
1459                         return err;
1460
1461         }
1462         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1463                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1464                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1465                         return -EOPNOTSUPP;
1466                 }
1467                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1468                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1469                                sizeof(key->ct_zone));
1470         }
1471         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1472                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1473                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1474                         return -EOPNOTSUPP;
1475                 }
1476                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1477                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1478                                sizeof(key->ct_mark));
1479         }
1480         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1481                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1482                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1483                         return -EOPNOTSUPP;
1484                 }
1485                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1486                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1487                                sizeof(key->ct_labels));
1488         }
1489
1490         return 0;
1491 }
1492
1493 static int fl_set_key(struct net *net, struct nlattr **tb,
1494                       struct fl_flow_key *key, struct fl_flow_key *mask,
1495                       struct netlink_ext_ack *extack)
1496 {
1497         __be16 ethertype;
1498         int ret = 0;
1499
1500         if (tb[TCA_FLOWER_INDEV]) {
1501                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1502                 if (err < 0)
1503                         return err;
1504                 key->meta.ingress_ifindex = err;
1505                 mask->meta.ingress_ifindex = 0xffffffff;
1506         }
1507
1508         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1509                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1510                        sizeof(key->eth.dst));
1511         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1512                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1513                        sizeof(key->eth.src));
1514
1515         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1516                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1517
1518                 if (eth_type_vlan(ethertype)) {
1519                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1520                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
1521                                         &mask->vlan);
1522
1523                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1524                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1525                                 if (eth_type_vlan(ethertype)) {
1526                                         fl_set_key_vlan(tb, ethertype,
1527                                                         TCA_FLOWER_KEY_CVLAN_ID,
1528                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1529                                                         &key->cvlan, &mask->cvlan);
1530                                         fl_set_key_val(tb, &key->basic.n_proto,
1531                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1532                                                        &mask->basic.n_proto,
1533                                                        TCA_FLOWER_UNSPEC,
1534                                                        sizeof(key->basic.n_proto));
1535                                 } else {
1536                                         key->basic.n_proto = ethertype;
1537                                         mask->basic.n_proto = cpu_to_be16(~0);
1538                                 }
1539                         }
1540                 } else {
1541                         key->basic.n_proto = ethertype;
1542                         mask->basic.n_proto = cpu_to_be16(~0);
1543                 }
1544         }
1545
1546         if (key->basic.n_proto == htons(ETH_P_IP) ||
1547             key->basic.n_proto == htons(ETH_P_IPV6)) {
1548                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1549                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1550                                sizeof(key->basic.ip_proto));
1551                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1552         }
1553
1554         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1555                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1556                 mask->control.addr_type = ~0;
1557                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1558                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1559                                sizeof(key->ipv4.src));
1560                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1561                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1562                                sizeof(key->ipv4.dst));
1563         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1564                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1565                 mask->control.addr_type = ~0;
1566                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1567                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1568                                sizeof(key->ipv6.src));
1569                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1570                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1571                                sizeof(key->ipv6.dst));
1572         }
1573
1574         if (key->basic.ip_proto == IPPROTO_TCP) {
1575                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1576                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1577                                sizeof(key->tp.src));
1578                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1579                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1580                                sizeof(key->tp.dst));
1581                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1582                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1583                                sizeof(key->tcp.flags));
1584         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1585                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1586                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1587                                sizeof(key->tp.src));
1588                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1589                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1590                                sizeof(key->tp.dst));
1591         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1592                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1593                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1594                                sizeof(key->tp.src));
1595                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1596                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1597                                sizeof(key->tp.dst));
1598         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1599                    key->basic.ip_proto == IPPROTO_ICMP) {
1600                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1601                                &mask->icmp.type,
1602                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1603                                sizeof(key->icmp.type));
1604                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1605                                &mask->icmp.code,
1606                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1607                                sizeof(key->icmp.code));
1608         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1609                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1610                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1611                                &mask->icmp.type,
1612                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1613                                sizeof(key->icmp.type));
1614                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1615                                &mask->icmp.code,
1616                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1617                                sizeof(key->icmp.code));
1618         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1619                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1620                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1621                 if (ret)
1622                         return ret;
1623         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1624                    key->basic.n_proto == htons(ETH_P_RARP)) {
1625                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1626                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1627                                sizeof(key->arp.sip));
1628                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1629                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1630                                sizeof(key->arp.tip));
1631                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1632                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1633                                sizeof(key->arp.op));
1634                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1635                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1636                                sizeof(key->arp.sha));
1637                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1638                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1639                                sizeof(key->arp.tha));
1640         }
1641
1642         if (key->basic.ip_proto == IPPROTO_TCP ||
1643             key->basic.ip_proto == IPPROTO_UDP ||
1644             key->basic.ip_proto == IPPROTO_SCTP) {
1645                 ret = fl_set_key_port_range(tb, key, mask, extack);
1646                 if (ret)
1647                         return ret;
1648         }
1649
1650         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1651             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1652                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1653                 mask->enc_control.addr_type = ~0;
1654                 fl_set_key_val(tb, &key->enc_ipv4.src,
1655                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1656                                &mask->enc_ipv4.src,
1657                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1658                                sizeof(key->enc_ipv4.src));
1659                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1660                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1661                                &mask->enc_ipv4.dst,
1662                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1663                                sizeof(key->enc_ipv4.dst));
1664         }
1665
1666         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1667             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1668                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1669                 mask->enc_control.addr_type = ~0;
1670                 fl_set_key_val(tb, &key->enc_ipv6.src,
1671                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1672                                &mask->enc_ipv6.src,
1673                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1674                                sizeof(key->enc_ipv6.src));
1675                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1676                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1677                                &mask->enc_ipv6.dst,
1678                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1679                                sizeof(key->enc_ipv6.dst));
1680         }
1681
1682         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1683                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1684                        sizeof(key->enc_key_id.keyid));
1685
1686         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1687                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1688                        sizeof(key->enc_tp.src));
1689
1690         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1691                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1692                        sizeof(key->enc_tp.dst));
1693
1694         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1695
1696         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1697                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1698                        sizeof(key->hash.hash));
1699
1700         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1701                 ret = fl_set_enc_opt(tb, key, mask, extack);
1702                 if (ret)
1703                         return ret;
1704         }
1705
1706         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1707         if (ret)
1708                 return ret;
1709
1710         if (tb[TCA_FLOWER_KEY_FLAGS])
1711                 ret = fl_set_key_flags(tb, &key->control.flags,
1712                                        &mask->control.flags, extack);
1713
1714         return ret;
1715 }
1716
1717 static void fl_mask_copy(struct fl_flow_mask *dst,
1718                          struct fl_flow_mask *src)
1719 {
1720         const void *psrc = fl_key_get_start(&src->key, src);
1721         void *pdst = fl_key_get_start(&dst->key, src);
1722
1723         memcpy(pdst, psrc, fl_mask_range(src));
1724         dst->range = src->range;
1725 }
1726
1727 static const struct rhashtable_params fl_ht_params = {
1728         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1729         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1730         .automatic_shrinking = true,
1731 };
1732
1733 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1734 {
1735         mask->filter_ht_params = fl_ht_params;
1736         mask->filter_ht_params.key_len = fl_mask_range(mask);
1737         mask->filter_ht_params.key_offset += mask->range.start;
1738
1739         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1740 }
1741
1742 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1743 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1744
1745 #define FL_KEY_IS_MASKED(mask, member)                                          \
1746         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1747                    0, FL_KEY_MEMBER_SIZE(member))                               \
1748
1749 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1750         do {                                                                    \
1751                 keys[cnt].key_id = id;                                          \
1752                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1753                 cnt++;                                                          \
1754         } while(0);
1755
1756 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1757         do {                                                                    \
1758                 if (FL_KEY_IS_MASKED(mask, member))                             \
1759                         FL_KEY_SET(keys, cnt, id, member);                      \
1760         } while(0);
1761
1762 static void fl_init_dissector(struct flow_dissector *dissector,
1763                               struct fl_flow_key *mask)
1764 {
1765         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1766         size_t cnt = 0;
1767
1768         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1769                              FLOW_DISSECTOR_KEY_META, meta);
1770         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1771         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1772         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1773                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1774         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1775                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1776         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1777                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1778         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1779                              FLOW_DISSECTOR_KEY_PORTS, tp);
1780         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1781                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1782         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1783                              FLOW_DISSECTOR_KEY_IP, ip);
1784         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1785                              FLOW_DISSECTOR_KEY_TCP, tcp);
1786         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1787                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1788         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1789                              FLOW_DISSECTOR_KEY_ARP, arp);
1790         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1791                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1792         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1793                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1794         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1795                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1796         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1797                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1798         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1799                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1800         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1801                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1802         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1803             FL_KEY_IS_MASKED(mask, enc_ipv6))
1804                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1805                            enc_control);
1806         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1807                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1808         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1809                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1810         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1811                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1812         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1813                              FLOW_DISSECTOR_KEY_CT, ct);
1814         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1815                              FLOW_DISSECTOR_KEY_HASH, hash);
1816
1817         skb_flow_dissector_init(dissector, keys, cnt);
1818 }
1819
1820 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1821                                                struct fl_flow_mask *mask)
1822 {
1823         struct fl_flow_mask *newmask;
1824         int err;
1825
1826         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1827         if (!newmask)
1828                 return ERR_PTR(-ENOMEM);
1829
1830         fl_mask_copy(newmask, mask);
1831
1832         if ((newmask->key.tp_range.tp_min.dst &&
1833              newmask->key.tp_range.tp_max.dst) ||
1834             (newmask->key.tp_range.tp_min.src &&
1835              newmask->key.tp_range.tp_max.src))
1836                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1837
1838         err = fl_init_mask_hashtable(newmask);
1839         if (err)
1840                 goto errout_free;
1841
1842         fl_init_dissector(&newmask->dissector, &newmask->key);
1843
1844         INIT_LIST_HEAD_RCU(&newmask->filters);
1845
1846         refcount_set(&newmask->refcnt, 1);
1847         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1848                                       &newmask->ht_node, mask_ht_params);
1849         if (err)
1850                 goto errout_destroy;
1851
1852         spin_lock(&head->masks_lock);
1853         list_add_tail_rcu(&newmask->list, &head->masks);
1854         spin_unlock(&head->masks_lock);
1855
1856         return newmask;
1857
1858 errout_destroy:
1859         rhashtable_destroy(&newmask->ht);
1860 errout_free:
1861         kfree(newmask);
1862
1863         return ERR_PTR(err);
1864 }
1865
1866 static int fl_check_assign_mask(struct cls_fl_head *head,
1867                                 struct cls_fl_filter *fnew,
1868                                 struct cls_fl_filter *fold,
1869                                 struct fl_flow_mask *mask)
1870 {
1871         struct fl_flow_mask *newmask;
1872         int ret = 0;
1873
1874         rcu_read_lock();
1875
1876         /* Insert mask as temporary node to prevent concurrent creation of mask
1877          * with same key. Any concurrent lookups with same key will return
1878          * -EAGAIN because mask's refcnt is zero.
1879          */
1880         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1881                                                        &mask->ht_node,
1882                                                        mask_ht_params);
1883         if (!fnew->mask) {
1884                 rcu_read_unlock();
1885
1886                 if (fold) {
1887                         ret = -EINVAL;
1888                         goto errout_cleanup;
1889                 }
1890
1891                 newmask = fl_create_new_mask(head, mask);
1892                 if (IS_ERR(newmask)) {
1893                         ret = PTR_ERR(newmask);
1894                         goto errout_cleanup;
1895                 }
1896
1897                 fnew->mask = newmask;
1898                 return 0;
1899         } else if (IS_ERR(fnew->mask)) {
1900                 ret = PTR_ERR(fnew->mask);
1901         } else if (fold && fold->mask != fnew->mask) {
1902                 ret = -EINVAL;
1903         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1904                 /* Mask was deleted concurrently, try again */
1905                 ret = -EAGAIN;
1906         }
1907         rcu_read_unlock();
1908         return ret;
1909
1910 errout_cleanup:
1911         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1912                                mask_ht_params);
1913         return ret;
1914 }
1915
1916 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1917                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1918                         unsigned long base, struct nlattr **tb,
1919                         struct nlattr *est,
1920                         struct fl_flow_tmplt *tmplt, u32 flags,
1921                         struct netlink_ext_ack *extack)
1922 {
1923         int err;
1924
1925         err = tcf_exts_validate(net, tp, tb, est, &f->exts, flags, extack);
1926         if (err < 0)
1927                 return err;
1928
1929         if (tb[TCA_FLOWER_CLASSID]) {
1930                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1931                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
1932                         rtnl_lock();
1933                 tcf_bind_filter(tp, &f->res, base);
1934                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
1935                         rtnl_unlock();
1936         }
1937
1938         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1939         if (err)
1940                 return err;
1941
1942         fl_mask_update_range(mask);
1943         fl_set_masked_key(&f->mkey, &f->key, mask);
1944
1945         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1946                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1947                 return -EINVAL;
1948         }
1949
1950         return 0;
1951 }
1952
1953 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
1954                                struct cls_fl_filter *fold,
1955                                bool *in_ht)
1956 {
1957         struct fl_flow_mask *mask = fnew->mask;
1958         int err;
1959
1960         err = rhashtable_lookup_insert_fast(&mask->ht,
1961                                             &fnew->ht_node,
1962                                             mask->filter_ht_params);
1963         if (err) {
1964                 *in_ht = false;
1965                 /* It is okay if filter with same key exists when
1966                  * overwriting.
1967                  */
1968                 return fold && err == -EEXIST ? 0 : err;
1969         }
1970
1971         *in_ht = true;
1972         return 0;
1973 }
1974
1975 static int fl_change(struct net *net, struct sk_buff *in_skb,
1976                      struct tcf_proto *tp, unsigned long base,
1977                      u32 handle, struct nlattr **tca,
1978                      void **arg, u32 flags,
1979                      struct netlink_ext_ack *extack)
1980 {
1981         struct cls_fl_head *head = fl_head_dereference(tp);
1982         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
1983         struct cls_fl_filter *fold = *arg;
1984         struct cls_fl_filter *fnew;
1985         struct fl_flow_mask *mask;
1986         struct nlattr **tb;
1987         bool in_ht;
1988         int err;
1989
1990         if (!tca[TCA_OPTIONS]) {
1991                 err = -EINVAL;
1992                 goto errout_fold;
1993         }
1994
1995         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
1996         if (!mask) {
1997                 err = -ENOBUFS;
1998                 goto errout_fold;
1999         }
2000
2001         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2002         if (!tb) {
2003                 err = -ENOBUFS;
2004                 goto errout_mask_alloc;
2005         }
2006
2007         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2008                                           tca[TCA_OPTIONS], fl_policy, NULL);
2009         if (err < 0)
2010                 goto errout_tb;
2011
2012         if (fold && handle && fold->handle != handle) {
2013                 err = -EINVAL;
2014                 goto errout_tb;
2015         }
2016
2017         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2018         if (!fnew) {
2019                 err = -ENOBUFS;
2020                 goto errout_tb;
2021         }
2022         INIT_LIST_HEAD(&fnew->hw_list);
2023         refcount_set(&fnew->refcnt, 1);
2024
2025         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
2026         if (err < 0)
2027                 goto errout;
2028
2029         if (tb[TCA_FLOWER_FLAGS]) {
2030                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2031
2032                 if (!tc_flags_valid(fnew->flags)) {
2033                         err = -EINVAL;
2034                         goto errout;
2035                 }
2036         }
2037
2038         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2039                            tp->chain->tmplt_priv, flags, extack);
2040         if (err)
2041                 goto errout;
2042
2043         err = fl_check_assign_mask(head, fnew, fold, mask);
2044         if (err)
2045                 goto errout;
2046
2047         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2048         if (err)
2049                 goto errout_mask;
2050
2051         if (!tc_skip_hw(fnew->flags)) {
2052                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2053                 if (err)
2054                         goto errout_ht;
2055         }
2056
2057         if (!tc_in_hw(fnew->flags))
2058                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2059
2060         spin_lock(&tp->lock);
2061
2062         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2063          * proto again or create new one, if necessary.
2064          */
2065         if (tp->deleting) {
2066                 err = -EAGAIN;
2067                 goto errout_hw;
2068         }
2069
2070         if (fold) {
2071                 /* Fold filter was deleted concurrently. Retry lookup. */
2072                 if (fold->deleted) {
2073                         err = -EAGAIN;
2074                         goto errout_hw;
2075                 }
2076
2077                 fnew->handle = handle;
2078
2079                 if (!in_ht) {
2080                         struct rhashtable_params params =
2081                                 fnew->mask->filter_ht_params;
2082
2083                         err = rhashtable_insert_fast(&fnew->mask->ht,
2084                                                      &fnew->ht_node,
2085                                                      params);
2086                         if (err)
2087                                 goto errout_hw;
2088                         in_ht = true;
2089                 }
2090
2091                 refcount_inc(&fnew->refcnt);
2092                 rhashtable_remove_fast(&fold->mask->ht,
2093                                        &fold->ht_node,
2094                                        fold->mask->filter_ht_params);
2095                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2096                 list_replace_rcu(&fold->list, &fnew->list);
2097                 fold->deleted = true;
2098
2099                 spin_unlock(&tp->lock);
2100
2101                 fl_mask_put(head, fold->mask);
2102                 if (!tc_skip_hw(fold->flags))
2103                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2104                 tcf_unbind_filter(tp, &fold->res);
2105                 /* Caller holds reference to fold, so refcnt is always > 0
2106                  * after this.
2107                  */
2108                 refcount_dec(&fold->refcnt);
2109                 __fl_put(fold);
2110         } else {
2111                 if (handle) {
2112                         /* user specifies a handle and it doesn't exist */
2113                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2114                                             handle, GFP_ATOMIC);
2115
2116                         /* Filter with specified handle was concurrently
2117                          * inserted after initial check in cls_api. This is not
2118                          * necessarily an error if NLM_F_EXCL is not set in
2119                          * message flags. Returning EAGAIN will cause cls_api to
2120                          * try to update concurrently inserted rule.
2121                          */
2122                         if (err == -ENOSPC)
2123                                 err = -EAGAIN;
2124                 } else {
2125                         handle = 1;
2126                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2127                                             INT_MAX, GFP_ATOMIC);
2128                 }
2129                 if (err)
2130                         goto errout_hw;
2131
2132                 refcount_inc(&fnew->refcnt);
2133                 fnew->handle = handle;
2134                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2135                 spin_unlock(&tp->lock);
2136         }
2137
2138         *arg = fnew;
2139
2140         kfree(tb);
2141         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2142         return 0;
2143
2144 errout_ht:
2145         spin_lock(&tp->lock);
2146 errout_hw:
2147         fnew->deleted = true;
2148         spin_unlock(&tp->lock);
2149         if (!tc_skip_hw(fnew->flags))
2150                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2151         if (in_ht)
2152                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2153                                        fnew->mask->filter_ht_params);
2154 errout_mask:
2155         fl_mask_put(head, fnew->mask);
2156 errout:
2157         __fl_put(fnew);
2158 errout_tb:
2159         kfree(tb);
2160 errout_mask_alloc:
2161         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2162 errout_fold:
2163         if (fold)
2164                 __fl_put(fold);
2165         return err;
2166 }
2167
2168 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2169                      bool rtnl_held, struct netlink_ext_ack *extack)
2170 {
2171         struct cls_fl_head *head = fl_head_dereference(tp);
2172         struct cls_fl_filter *f = arg;
2173         bool last_on_mask;
2174         int err = 0;
2175
2176         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2177         *last = list_empty(&head->masks);
2178         __fl_put(f);
2179
2180         return err;
2181 }
2182
2183 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2184                     bool rtnl_held)
2185 {
2186         struct cls_fl_head *head = fl_head_dereference(tp);
2187         unsigned long id = arg->cookie, tmp;
2188         struct cls_fl_filter *f;
2189
2190         arg->count = arg->skip;
2191
2192         rcu_read_lock();
2193         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2194                 /* don't return filters that are being deleted */
2195                 if (!refcount_inc_not_zero(&f->refcnt))
2196                         continue;
2197                 rcu_read_unlock();
2198
2199                 if (arg->fn(tp, f, arg) < 0) {
2200                         __fl_put(f);
2201                         arg->stop = 1;
2202                         rcu_read_lock();
2203                         break;
2204                 }
2205                 __fl_put(f);
2206                 arg->count++;
2207                 rcu_read_lock();
2208         }
2209         rcu_read_unlock();
2210         arg->cookie = id;
2211 }
2212
2213 static struct cls_fl_filter *
2214 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2215 {
2216         struct cls_fl_head *head = fl_head_dereference(tp);
2217
2218         spin_lock(&tp->lock);
2219         if (list_empty(&head->hw_filters)) {
2220                 spin_unlock(&tp->lock);
2221                 return NULL;
2222         }
2223
2224         if (!f)
2225                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2226                                hw_list);
2227         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2228                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2229                         spin_unlock(&tp->lock);
2230                         return f;
2231                 }
2232         }
2233
2234         spin_unlock(&tp->lock);
2235         return NULL;
2236 }
2237
2238 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2239                         void *cb_priv, struct netlink_ext_ack *extack)
2240 {
2241         struct tcf_block *block = tp->chain->block;
2242         struct flow_cls_offload cls_flower = {};
2243         struct cls_fl_filter *f = NULL;
2244         int err;
2245
2246         /* hw_filters list can only be changed by hw offload functions after
2247          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2248          * iterating it.
2249          */
2250         ASSERT_RTNL();
2251
2252         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2253                 cls_flower.rule =
2254                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2255                 if (!cls_flower.rule) {
2256                         __fl_put(f);
2257                         return -ENOMEM;
2258                 }
2259
2260                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2261                                            extack);
2262                 cls_flower.command = add ?
2263                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2264                 cls_flower.cookie = (unsigned long)f;
2265                 cls_flower.rule->match.dissector = &f->mask->dissector;
2266                 cls_flower.rule->match.mask = &f->mask->key;
2267                 cls_flower.rule->match.key = &f->mkey;
2268
2269                 err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
2270                 if (err) {
2271                         kfree(cls_flower.rule);
2272                         if (tc_skip_sw(f->flags)) {
2273                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
2274                                 __fl_put(f);
2275                                 return err;
2276                         }
2277                         goto next_flow;
2278                 }
2279
2280                 cls_flower.classid = f->res.classid;
2281
2282                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2283                                             TC_SETUP_CLSFLOWER, &cls_flower,
2284                                             cb_priv, &f->flags,
2285                                             &f->in_hw_count);
2286                 tc_cleanup_flow_action(&cls_flower.rule->action);
2287                 kfree(cls_flower.rule);
2288
2289                 if (err) {
2290                         __fl_put(f);
2291                         return err;
2292                 }
2293 next_flow:
2294                 __fl_put(f);
2295         }
2296
2297         return 0;
2298 }
2299
2300 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2301 {
2302         struct flow_cls_offload *cls_flower = type_data;
2303         struct cls_fl_filter *f =
2304                 (struct cls_fl_filter *) cls_flower->cookie;
2305         struct cls_fl_head *head = fl_head_dereference(tp);
2306
2307         spin_lock(&tp->lock);
2308         list_add(&f->hw_list, &head->hw_filters);
2309         spin_unlock(&tp->lock);
2310 }
2311
2312 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2313 {
2314         struct flow_cls_offload *cls_flower = type_data;
2315         struct cls_fl_filter *f =
2316                 (struct cls_fl_filter *) cls_flower->cookie;
2317
2318         spin_lock(&tp->lock);
2319         if (!list_empty(&f->hw_list))
2320                 list_del_init(&f->hw_list);
2321         spin_unlock(&tp->lock);
2322 }
2323
2324 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2325                               struct fl_flow_tmplt *tmplt)
2326 {
2327         struct flow_cls_offload cls_flower = {};
2328         struct tcf_block *block = chain->block;
2329
2330         cls_flower.rule = flow_rule_alloc(0);
2331         if (!cls_flower.rule)
2332                 return -ENOMEM;
2333
2334         cls_flower.common.chain_index = chain->index;
2335         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2336         cls_flower.cookie = (unsigned long) tmplt;
2337         cls_flower.rule->match.dissector = &tmplt->dissector;
2338         cls_flower.rule->match.mask = &tmplt->mask;
2339         cls_flower.rule->match.key = &tmplt->dummy_key;
2340
2341         /* We don't care if driver (any of them) fails to handle this
2342          * call. It serves just as a hint for it.
2343          */
2344         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2345         kfree(cls_flower.rule);
2346
2347         return 0;
2348 }
2349
2350 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2351                                 struct fl_flow_tmplt *tmplt)
2352 {
2353         struct flow_cls_offload cls_flower = {};
2354         struct tcf_block *block = chain->block;
2355
2356         cls_flower.common.chain_index = chain->index;
2357         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2358         cls_flower.cookie = (unsigned long) tmplt;
2359
2360         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2361 }
2362
2363 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2364                              struct nlattr **tca,
2365                              struct netlink_ext_ack *extack)
2366 {
2367         struct fl_flow_tmplt *tmplt;
2368         struct nlattr **tb;
2369         int err;
2370
2371         if (!tca[TCA_OPTIONS])
2372                 return ERR_PTR(-EINVAL);
2373
2374         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2375         if (!tb)
2376                 return ERR_PTR(-ENOBUFS);
2377         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2378                                           tca[TCA_OPTIONS], fl_policy, NULL);
2379         if (err)
2380                 goto errout_tb;
2381
2382         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2383         if (!tmplt) {
2384                 err = -ENOMEM;
2385                 goto errout_tb;
2386         }
2387         tmplt->chain = chain;
2388         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2389         if (err)
2390                 goto errout_tmplt;
2391
2392         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2393
2394         err = fl_hw_create_tmplt(chain, tmplt);
2395         if (err)
2396                 goto errout_tmplt;
2397
2398         kfree(tb);
2399         return tmplt;
2400
2401 errout_tmplt:
2402         kfree(tmplt);
2403 errout_tb:
2404         kfree(tb);
2405         return ERR_PTR(err);
2406 }
2407
2408 static void fl_tmplt_destroy(void *tmplt_priv)
2409 {
2410         struct fl_flow_tmplt *tmplt = tmplt_priv;
2411
2412         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2413         kfree(tmplt);
2414 }
2415
2416 static int fl_dump_key_val(struct sk_buff *skb,
2417                            void *val, int val_type,
2418                            void *mask, int mask_type, int len)
2419 {
2420         int err;
2421
2422         if (!memchr_inv(mask, 0, len))
2423                 return 0;
2424         err = nla_put(skb, val_type, len, val);
2425         if (err)
2426                 return err;
2427         if (mask_type != TCA_FLOWER_UNSPEC) {
2428                 err = nla_put(skb, mask_type, len, mask);
2429                 if (err)
2430                         return err;
2431         }
2432         return 0;
2433 }
2434
2435 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2436                                   struct fl_flow_key *mask)
2437 {
2438         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2439                             TCA_FLOWER_KEY_PORT_DST_MIN,
2440                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2441                             sizeof(key->tp_range.tp_min.dst)) ||
2442             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2443                             TCA_FLOWER_KEY_PORT_DST_MAX,
2444                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2445                             sizeof(key->tp_range.tp_max.dst)) ||
2446             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2447                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2448                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2449                             sizeof(key->tp_range.tp_min.src)) ||
2450             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2451                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2452                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2453                             sizeof(key->tp_range.tp_max.src)))
2454                 return -1;
2455
2456         return 0;
2457 }
2458
2459 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2460                                     struct flow_dissector_key_mpls *mpls_key,
2461                                     struct flow_dissector_key_mpls *mpls_mask,
2462                                     u8 lse_index)
2463 {
2464         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2465         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2466         int err;
2467
2468         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2469                          lse_index + 1);
2470         if (err)
2471                 return err;
2472
2473         if (lse_mask->mpls_ttl) {
2474                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2475                                  lse_key->mpls_ttl);
2476                 if (err)
2477                         return err;
2478         }
2479         if (lse_mask->mpls_bos) {
2480                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2481                                  lse_key->mpls_bos);
2482                 if (err)
2483                         return err;
2484         }
2485         if (lse_mask->mpls_tc) {
2486                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2487                                  lse_key->mpls_tc);
2488                 if (err)
2489                         return err;
2490         }
2491         if (lse_mask->mpls_label) {
2492                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2493                                   lse_key->mpls_label);
2494                 if (err)
2495                         return err;
2496         }
2497
2498         return 0;
2499 }
2500
2501 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2502                                  struct flow_dissector_key_mpls *mpls_key,
2503                                  struct flow_dissector_key_mpls *mpls_mask)
2504 {
2505         struct nlattr *opts;
2506         struct nlattr *lse;
2507         u8 lse_index;
2508         int err;
2509
2510         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2511         if (!opts)
2512                 return -EMSGSIZE;
2513
2514         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2515                 if (!(mpls_mask->used_lses & 1 << lse_index))
2516                         continue;
2517
2518                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2519                 if (!lse) {
2520                         err = -EMSGSIZE;
2521                         goto err_opts;
2522                 }
2523
2524                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2525                                                lse_index);
2526                 if (err)
2527                         goto err_opts_lse;
2528                 nla_nest_end(skb, lse);
2529         }
2530         nla_nest_end(skb, opts);
2531
2532         return 0;
2533
2534 err_opts_lse:
2535         nla_nest_cancel(skb, lse);
2536 err_opts:
2537         nla_nest_cancel(skb, opts);
2538
2539         return err;
2540 }
2541
2542 static int fl_dump_key_mpls(struct sk_buff *skb,
2543                             struct flow_dissector_key_mpls *mpls_key,
2544                             struct flow_dissector_key_mpls *mpls_mask)
2545 {
2546         struct flow_dissector_mpls_lse *lse_mask;
2547         struct flow_dissector_mpls_lse *lse_key;
2548         int err;
2549
2550         if (!mpls_mask->used_lses)
2551                 return 0;
2552
2553         lse_mask = &mpls_mask->ls[0];
2554         lse_key = &mpls_key->ls[0];
2555
2556         /* For backward compatibility, don't use the MPLS nested attributes if
2557          * the rule can be expressed using the old attributes.
2558          */
2559         if (mpls_mask->used_lses & ~1 ||
2560             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2561              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2562                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2563
2564         if (lse_mask->mpls_ttl) {
2565                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2566                                  lse_key->mpls_ttl);
2567                 if (err)
2568                         return err;
2569         }
2570         if (lse_mask->mpls_tc) {
2571                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2572                                  lse_key->mpls_tc);
2573                 if (err)
2574                         return err;
2575         }
2576         if (lse_mask->mpls_label) {
2577                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2578                                   lse_key->mpls_label);
2579                 if (err)
2580                         return err;
2581         }
2582         if (lse_mask->mpls_bos) {
2583                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2584                                  lse_key->mpls_bos);
2585                 if (err)
2586                         return err;
2587         }
2588         return 0;
2589 }
2590
2591 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2592                           struct flow_dissector_key_ip *key,
2593                           struct flow_dissector_key_ip *mask)
2594 {
2595         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2596         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2597         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2598         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2599
2600         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2601             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2602                 return -1;
2603
2604         return 0;
2605 }
2606
2607 static int fl_dump_key_vlan(struct sk_buff *skb,
2608                             int vlan_id_key, int vlan_prio_key,
2609                             struct flow_dissector_key_vlan *vlan_key,
2610                             struct flow_dissector_key_vlan *vlan_mask)
2611 {
2612         int err;
2613
2614         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2615                 return 0;
2616         if (vlan_mask->vlan_id) {
2617                 err = nla_put_u16(skb, vlan_id_key,
2618                                   vlan_key->vlan_id);
2619                 if (err)
2620                         return err;
2621         }
2622         if (vlan_mask->vlan_priority) {
2623                 err = nla_put_u8(skb, vlan_prio_key,
2624                                  vlan_key->vlan_priority);
2625                 if (err)
2626                         return err;
2627         }
2628         return 0;
2629 }
2630
2631 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2632                             u32 *flower_key, u32 *flower_mask,
2633                             u32 flower_flag_bit, u32 dissector_flag_bit)
2634 {
2635         if (dissector_mask & dissector_flag_bit) {
2636                 *flower_mask |= flower_flag_bit;
2637                 if (dissector_key & dissector_flag_bit)
2638                         *flower_key |= flower_flag_bit;
2639         }
2640 }
2641
2642 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2643 {
2644         u32 key, mask;
2645         __be32 _key, _mask;
2646         int err;
2647
2648         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2649                 return 0;
2650
2651         key = 0;
2652         mask = 0;
2653
2654         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2655                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2656         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2657                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2658                         FLOW_DIS_FIRST_FRAG);
2659
2660         _key = cpu_to_be32(key);
2661         _mask = cpu_to_be32(mask);
2662
2663         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2664         if (err)
2665                 return err;
2666
2667         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2668 }
2669
2670 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2671                                   struct flow_dissector_key_enc_opts *enc_opts)
2672 {
2673         struct geneve_opt *opt;
2674         struct nlattr *nest;
2675         int opt_off = 0;
2676
2677         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2678         if (!nest)
2679                 goto nla_put_failure;
2680
2681         while (enc_opts->len > opt_off) {
2682                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2683
2684                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2685                                  opt->opt_class))
2686                         goto nla_put_failure;
2687                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2688                                opt->type))
2689                         goto nla_put_failure;
2690                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2691                             opt->length * 4, opt->opt_data))
2692                         goto nla_put_failure;
2693
2694                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2695         }
2696         nla_nest_end(skb, nest);
2697         return 0;
2698
2699 nla_put_failure:
2700         nla_nest_cancel(skb, nest);
2701         return -EMSGSIZE;
2702 }
2703
2704 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2705                                  struct flow_dissector_key_enc_opts *enc_opts)
2706 {
2707         struct vxlan_metadata *md;
2708         struct nlattr *nest;
2709
2710         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2711         if (!nest)
2712                 goto nla_put_failure;
2713
2714         md = (struct vxlan_metadata *)&enc_opts->data[0];
2715         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2716                 goto nla_put_failure;
2717
2718         nla_nest_end(skb, nest);
2719         return 0;
2720
2721 nla_put_failure:
2722         nla_nest_cancel(skb, nest);
2723         return -EMSGSIZE;
2724 }
2725
2726 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2727                                   struct flow_dissector_key_enc_opts *enc_opts)
2728 {
2729         struct erspan_metadata *md;
2730         struct nlattr *nest;
2731
2732         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2733         if (!nest)
2734                 goto nla_put_failure;
2735
2736         md = (struct erspan_metadata *)&enc_opts->data[0];
2737         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2738                 goto nla_put_failure;
2739
2740         if (md->version == 1 &&
2741             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2742                 goto nla_put_failure;
2743
2744         if (md->version == 2 &&
2745             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2746                         md->u.md2.dir) ||
2747              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2748                         get_hwid(&md->u.md2))))
2749                 goto nla_put_failure;
2750
2751         nla_nest_end(skb, nest);
2752         return 0;
2753
2754 nla_put_failure:
2755         nla_nest_cancel(skb, nest);
2756         return -EMSGSIZE;
2757 }
2758
2759 static int fl_dump_key_ct(struct sk_buff *skb,
2760                           struct flow_dissector_key_ct *key,
2761                           struct flow_dissector_key_ct *mask)
2762 {
2763         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2764             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2765                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2766                             sizeof(key->ct_state)))
2767                 goto nla_put_failure;
2768
2769         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2770             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2771                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2772                             sizeof(key->ct_zone)))
2773                 goto nla_put_failure;
2774
2775         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2776             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2777                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2778                             sizeof(key->ct_mark)))
2779                 goto nla_put_failure;
2780
2781         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2782             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2783                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2784                             sizeof(key->ct_labels)))
2785                 goto nla_put_failure;
2786
2787         return 0;
2788
2789 nla_put_failure:
2790         return -EMSGSIZE;
2791 }
2792
2793 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2794                                struct flow_dissector_key_enc_opts *enc_opts)
2795 {
2796         struct nlattr *nest;
2797         int err;
2798
2799         if (!enc_opts->len)
2800                 return 0;
2801
2802         nest = nla_nest_start_noflag(skb, enc_opt_type);
2803         if (!nest)
2804                 goto nla_put_failure;
2805
2806         switch (enc_opts->dst_opt_type) {
2807         case TUNNEL_GENEVE_OPT:
2808                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2809                 if (err)
2810                         goto nla_put_failure;
2811                 break;
2812         case TUNNEL_VXLAN_OPT:
2813                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
2814                 if (err)
2815                         goto nla_put_failure;
2816                 break;
2817         case TUNNEL_ERSPAN_OPT:
2818                 err = fl_dump_key_erspan_opt(skb, enc_opts);
2819                 if (err)
2820                         goto nla_put_failure;
2821                 break;
2822         default:
2823                 goto nla_put_failure;
2824         }
2825         nla_nest_end(skb, nest);
2826         return 0;
2827
2828 nla_put_failure:
2829         nla_nest_cancel(skb, nest);
2830         return -EMSGSIZE;
2831 }
2832
2833 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2834                                struct flow_dissector_key_enc_opts *key_opts,
2835                                struct flow_dissector_key_enc_opts *msk_opts)
2836 {
2837         int err;
2838
2839         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2840         if (err)
2841                 return err;
2842
2843         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2844 }
2845
2846 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2847                        struct fl_flow_key *key, struct fl_flow_key *mask)
2848 {
2849         if (mask->meta.ingress_ifindex) {
2850                 struct net_device *dev;
2851
2852                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
2853                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2854                         goto nla_put_failure;
2855         }
2856
2857         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2858                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2859                             sizeof(key->eth.dst)) ||
2860             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2861                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2862                             sizeof(key->eth.src)) ||
2863             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2864                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2865                             sizeof(key->basic.n_proto)))
2866                 goto nla_put_failure;
2867
2868         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2869                 goto nla_put_failure;
2870
2871         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2872                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2873                 goto nla_put_failure;
2874
2875         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2876                              TCA_FLOWER_KEY_CVLAN_PRIO,
2877                              &key->cvlan, &mask->cvlan) ||
2878             (mask->cvlan.vlan_tpid &&
2879              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2880                           key->cvlan.vlan_tpid)))
2881                 goto nla_put_failure;
2882
2883         if (mask->basic.n_proto) {
2884                 if (mask->cvlan.vlan_tpid) {
2885                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
2886                                          key->basic.n_proto))
2887                                 goto nla_put_failure;
2888                 } else if (mask->vlan.vlan_tpid) {
2889                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2890                                          key->basic.n_proto))
2891                                 goto nla_put_failure;
2892                 }
2893         }
2894
2895         if ((key->basic.n_proto == htons(ETH_P_IP) ||
2896              key->basic.n_proto == htons(ETH_P_IPV6)) &&
2897             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
2898                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
2899                             sizeof(key->basic.ip_proto)) ||
2900             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
2901                 goto nla_put_failure;
2902
2903         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2904             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
2905                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
2906                              sizeof(key->ipv4.src)) ||
2907              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
2908                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
2909                              sizeof(key->ipv4.dst))))
2910                 goto nla_put_failure;
2911         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2912                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
2913                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
2914                                   sizeof(key->ipv6.src)) ||
2915                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
2916                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
2917                                   sizeof(key->ipv6.dst))))
2918                 goto nla_put_failure;
2919
2920         if (key->basic.ip_proto == IPPROTO_TCP &&
2921             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
2922                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
2923                              sizeof(key->tp.src)) ||
2924              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
2925                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
2926                              sizeof(key->tp.dst)) ||
2927              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
2928                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
2929                              sizeof(key->tcp.flags))))
2930                 goto nla_put_failure;
2931         else if (key->basic.ip_proto == IPPROTO_UDP &&
2932                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
2933                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
2934                                   sizeof(key->tp.src)) ||
2935                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
2936                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
2937                                   sizeof(key->tp.dst))))
2938                 goto nla_put_failure;
2939         else if (key->basic.ip_proto == IPPROTO_SCTP &&
2940                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
2941                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
2942                                   sizeof(key->tp.src)) ||
2943                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
2944                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
2945                                   sizeof(key->tp.dst))))
2946                 goto nla_put_failure;
2947         else if (key->basic.n_proto == htons(ETH_P_IP) &&
2948                  key->basic.ip_proto == IPPROTO_ICMP &&
2949                  (fl_dump_key_val(skb, &key->icmp.type,
2950                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
2951                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
2952                                   sizeof(key->icmp.type)) ||
2953                   fl_dump_key_val(skb, &key->icmp.code,
2954                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
2955                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
2956                                   sizeof(key->icmp.code))))
2957                 goto nla_put_failure;
2958         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
2959                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
2960                  (fl_dump_key_val(skb, &key->icmp.type,
2961                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
2962                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
2963                                   sizeof(key->icmp.type)) ||
2964                   fl_dump_key_val(skb, &key->icmp.code,
2965                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
2966                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
2967                                   sizeof(key->icmp.code))))
2968                 goto nla_put_failure;
2969         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
2970                   key->basic.n_proto == htons(ETH_P_RARP)) &&
2971                  (fl_dump_key_val(skb, &key->arp.sip,
2972                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
2973                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
2974                                   sizeof(key->arp.sip)) ||
2975                   fl_dump_key_val(skb, &key->arp.tip,
2976                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
2977                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
2978                                   sizeof(key->arp.tip)) ||
2979                   fl_dump_key_val(skb, &key->arp.op,
2980                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
2981                                   TCA_FLOWER_KEY_ARP_OP_MASK,
2982                                   sizeof(key->arp.op)) ||
2983                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2984                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2985                                   sizeof(key->arp.sha)) ||
2986                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2987                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2988                                   sizeof(key->arp.tha))))
2989                 goto nla_put_failure;
2990
2991         if ((key->basic.ip_proto == IPPROTO_TCP ||
2992              key->basic.ip_proto == IPPROTO_UDP ||
2993              key->basic.ip_proto == IPPROTO_SCTP) &&
2994              fl_dump_key_port_range(skb, key, mask))
2995                 goto nla_put_failure;
2996
2997         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2998             (fl_dump_key_val(skb, &key->enc_ipv4.src,
2999                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3000                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3001                             sizeof(key->enc_ipv4.src)) ||
3002              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3003                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3004                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3005                              sizeof(key->enc_ipv4.dst))))
3006                 goto nla_put_failure;
3007         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3008                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3009                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3010                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3011                             sizeof(key->enc_ipv6.src)) ||
3012                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3013                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3014                                  &mask->enc_ipv6.dst,
3015                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3016                             sizeof(key->enc_ipv6.dst))))
3017                 goto nla_put_failure;
3018
3019         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3020                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3021                             sizeof(key->enc_key_id)) ||
3022             fl_dump_key_val(skb, &key->enc_tp.src,
3023                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3024                             &mask->enc_tp.src,
3025                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3026                             sizeof(key->enc_tp.src)) ||
3027             fl_dump_key_val(skb, &key->enc_tp.dst,
3028                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3029                             &mask->enc_tp.dst,
3030                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3031                             sizeof(key->enc_tp.dst)) ||
3032             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3033             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3034                 goto nla_put_failure;
3035
3036         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3037                 goto nla_put_failure;
3038
3039         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3040                 goto nla_put_failure;
3041
3042         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3043                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3044                              sizeof(key->hash.hash)))
3045                 goto nla_put_failure;
3046
3047         return 0;
3048
3049 nla_put_failure:
3050         return -EMSGSIZE;
3051 }
3052
3053 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3054                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3055 {
3056         struct cls_fl_filter *f = fh;
3057         struct nlattr *nest;
3058         struct fl_flow_key *key, *mask;
3059         bool skip_hw;
3060
3061         if (!f)
3062                 return skb->len;
3063
3064         t->tcm_handle = f->handle;
3065
3066         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3067         if (!nest)
3068                 goto nla_put_failure;
3069
3070         spin_lock(&tp->lock);
3071
3072         if (f->res.classid &&
3073             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3074                 goto nla_put_failure_locked;
3075
3076         key = &f->key;
3077         mask = &f->mask->key;
3078         skip_hw = tc_skip_hw(f->flags);
3079
3080         if (fl_dump_key(skb, net, key, mask))
3081                 goto nla_put_failure_locked;
3082
3083         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3084                 goto nla_put_failure_locked;
3085
3086         spin_unlock(&tp->lock);
3087
3088         if (!skip_hw)
3089                 fl_hw_update_stats(tp, f, rtnl_held);
3090
3091         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3092                 goto nla_put_failure;
3093
3094         if (tcf_exts_dump(skb, &f->exts))
3095                 goto nla_put_failure;
3096
3097         nla_nest_end(skb, nest);
3098
3099         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3100                 goto nla_put_failure;
3101
3102         return skb->len;
3103
3104 nla_put_failure_locked:
3105         spin_unlock(&tp->lock);
3106 nla_put_failure:
3107         nla_nest_cancel(skb, nest);
3108         return -1;
3109 }
3110
3111 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3112                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3113 {
3114         struct cls_fl_filter *f = fh;
3115         struct nlattr *nest;
3116         bool skip_hw;
3117
3118         if (!f)
3119                 return skb->len;
3120
3121         t->tcm_handle = f->handle;
3122
3123         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3124         if (!nest)
3125                 goto nla_put_failure;
3126
3127         spin_lock(&tp->lock);
3128
3129         skip_hw = tc_skip_hw(f->flags);
3130
3131         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3132                 goto nla_put_failure_locked;
3133
3134         spin_unlock(&tp->lock);
3135
3136         if (!skip_hw)
3137                 fl_hw_update_stats(tp, f, rtnl_held);
3138
3139         if (tcf_exts_terse_dump(skb, &f->exts))
3140                 goto nla_put_failure;
3141
3142         nla_nest_end(skb, nest);
3143
3144         return skb->len;
3145
3146 nla_put_failure_locked:
3147         spin_unlock(&tp->lock);
3148 nla_put_failure:
3149         nla_nest_cancel(skb, nest);
3150         return -1;
3151 }
3152
3153 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3154 {
3155         struct fl_flow_tmplt *tmplt = tmplt_priv;
3156         struct fl_flow_key *key, *mask;
3157         struct nlattr *nest;
3158
3159         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3160         if (!nest)
3161                 goto nla_put_failure;
3162
3163         key = &tmplt->dummy_key;
3164         mask = &tmplt->mask;
3165
3166         if (fl_dump_key(skb, net, key, mask))
3167                 goto nla_put_failure;
3168
3169         nla_nest_end(skb, nest);
3170
3171         return skb->len;
3172
3173 nla_put_failure:
3174         nla_nest_cancel(skb, nest);
3175         return -EMSGSIZE;
3176 }
3177
3178 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3179                           unsigned long base)
3180 {
3181         struct cls_fl_filter *f = fh;
3182
3183         if (f && f->res.classid == classid) {
3184                 if (cl)
3185                         __tcf_bind_filter(q, &f->res, base);
3186                 else
3187                         __tcf_unbind_filter(q, &f->res);
3188         }
3189 }
3190
3191 static bool fl_delete_empty(struct tcf_proto *tp)
3192 {
3193         struct cls_fl_head *head = fl_head_dereference(tp);
3194
3195         spin_lock(&tp->lock);
3196         tp->deleting = idr_is_empty(&head->handle_idr);
3197         spin_unlock(&tp->lock);
3198
3199         return tp->deleting;
3200 }
3201
3202 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3203         .kind           = "flower",
3204         .classify       = fl_classify,
3205         .init           = fl_init,
3206         .destroy        = fl_destroy,
3207         .get            = fl_get,
3208         .put            = fl_put,
3209         .change         = fl_change,
3210         .delete         = fl_delete,
3211         .delete_empty   = fl_delete_empty,
3212         .walk           = fl_walk,
3213         .reoffload      = fl_reoffload,
3214         .hw_add         = fl_hw_add,
3215         .hw_del         = fl_hw_del,
3216         .dump           = fl_dump,
3217         .terse_dump     = fl_terse_dump,
3218         .bind_class     = fl_bind_class,
3219         .tmplt_create   = fl_tmplt_create,
3220         .tmplt_destroy  = fl_tmplt_destroy,
3221         .tmplt_dump     = fl_tmplt_dump,
3222         .owner          = THIS_MODULE,
3223         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3224 };
3225
3226 static int __init cls_fl_init(void)
3227 {
3228         return register_tcf_proto_ops(&cls_fl_ops);
3229 }
3230
3231 static void __exit cls_fl_exit(void)
3232 {
3233         unregister_tcf_proto_ops(&cls_fl_ops);
3234 }
3235
3236 module_init(cls_fl_init);
3237 module_exit(cls_fl_exit);
3238
3239 MODULE_AUTHOR("Jiri Pirko <[email protected]>");
3240 MODULE_DESCRIPTION("Flower classifier");
3241 MODULE_LICENSE("GPL v2");