net/sched/cls_flower.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/cls_flower.c               Flower classifier
   4  *
   5  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/rhashtable.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/refcount.h>
  14
  15 #include <linux/if_ether.h>
  16 #include <linux/in6.h>
  17 #include <linux/ip.h>
  18 #include <linux/mpls.h>
  19 #include <linux/ppp_defs.h>
  20
  21 #include <net/sch_generic.h>
  22 #include <net/pkt_cls.h>
  23 #include <net/pkt_sched.h>
  24 #include <net/ip.h>
  25 #include <net/flow_dissector.h>
  26 #include <net/geneve.h>
  27 #include <net/vxlan.h>
  28 #include <net/erspan.h>
  29 #include <net/gtp.h>
  30 #include <net/tc_wrapper.h>
  31
  32 #include <net/dst.h>
  33 #include <net/dst_metadata.h>
  34
  35 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  36
  37 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  38                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  39 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  40                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  41
  42 struct fl_flow_key {
  43         struct flow_dissector_key_meta meta;
  44         struct flow_dissector_key_control control;
  45         struct flow_dissector_key_control enc_control;
  46         struct flow_dissector_key_basic basic;
  47         struct flow_dissector_key_eth_addrs eth;
  48         struct flow_dissector_key_vlan vlan;
  49         struct flow_dissector_key_vlan cvlan;
  50         union {
  51                 struct flow_dissector_key_ipv4_addrs ipv4;
  52                 struct flow_dissector_key_ipv6_addrs ipv6;
  53         };
  54         struct flow_dissector_key_ports tp;
  55         struct flow_dissector_key_icmp icmp;
  56         struct flow_dissector_key_arp arp;
  57         struct flow_dissector_key_keyid enc_key_id;
  58         union {
  59                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  60                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  61         };
  62         struct flow_dissector_key_ports enc_tp;
  63         struct flow_dissector_key_mpls mpls;
  64         struct flow_dissector_key_tcp tcp;
  65         struct flow_dissector_key_ip ip;
  66         struct flow_dissector_key_ip enc_ip;
  67         struct flow_dissector_key_enc_opts enc_opts;
  68         struct flow_dissector_key_ports_range tp_range;
  69         struct flow_dissector_key_ct ct;
  70         struct flow_dissector_key_hash hash;
  71         struct flow_dissector_key_num_of_vlans num_of_vlans;
  72         struct flow_dissector_key_pppoe pppoe;
  73         struct flow_dissector_key_l2tpv3 l2tpv3;
  74 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  75
  76 struct fl_flow_mask_range {
  77         unsigned short int start;
  78         unsigned short int end;
  79 };
  80
  81 struct fl_flow_mask {
  82         struct fl_flow_key key;
  83         struct fl_flow_mask_range range;
  84         u32 flags;
  85         struct rhash_head ht_node;
  86         struct rhashtable ht;
  87         struct rhashtable_params filter_ht_params;
  88         struct flow_dissector dissector;
  89         struct list_head filters;
  90         struct rcu_work rwork;
  91         struct list_head list;
  92         refcount_t refcnt;
  93 };
  94
  95 struct fl_flow_tmplt {
  96         struct fl_flow_key dummy_key;
  97         struct fl_flow_key mask;
  98         struct flow_dissector dissector;
  99         struct tcf_chain *chain;
 100 };
 101
 102 struct cls_fl_head {
 103         struct rhashtable ht;
 104         spinlock_t masks_lock; /* Protect masks list */
 105         struct list_head masks;
 106         struct list_head hw_filters;
 107         struct rcu_work rwork;
 108         struct idr handle_idr;
 109 };
 110
 111 struct cls_fl_filter {
 112         struct fl_flow_mask *mask;
 113         struct rhash_head ht_node;
 114         struct fl_flow_key mkey;
 115         struct tcf_exts exts;
 116         struct tcf_result res;
 117         struct fl_flow_key key;
 118         struct list_head list;
 119         struct list_head hw_list;
 120         u32 handle;
 121         u32 flags;
 122         u32 in_hw_count;
 123         struct rcu_work rwork;
 124         struct net_device *hw_dev;
 125         /* Flower classifier is unlocked, which means that its reference counter
 126          * can be changed concurrently without any kind of external
 127          * synchronization. Use atomic reference counter to be concurrency-safe.
 128          */
 129         refcount_t refcnt;
 130         bool deleted;
 131 };
 132
 133 static const struct rhashtable_params mask_ht_params = {
 134         .key_offset = offsetof(struct fl_flow_mask, key),
 135         .key_len = sizeof(struct fl_flow_key),
 136         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 137         .automatic_shrinking = true,
 138 };
 139
 140 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 141 {
 142         return mask->range.end - mask->range.start;
 143 }
 144
 145 static void fl_mask_update_range(struct fl_flow_mask *mask)
 146 {
 147         const u8 *bytes = (const u8 *) &mask->key;
 148         size_t size = sizeof(mask->key);
 149         size_t i, first = 0, last;
 150
 151         for (i = 0; i < size; i++) {
 152                 if (bytes[i]) {
 153                         first = i;
 154                         break;
 155                 }
 156         }
 157         last = first;
 158         for (i = size - 1; i != first; i--) {
 159                 if (bytes[i]) {
 160                         last = i;
 161                         break;
 162                 }
 163         }
 164         mask->range.start = rounddown(first, sizeof(long));
 165         mask->range.end = roundup(last + 1, sizeof(long));
 166 }
 167
 168 static void *fl_key_get_start(struct fl_flow_key *key,
 169                               const struct fl_flow_mask *mask)
 170 {
 171         return (u8 *) key + mask->range.start;
 172 }
 173
 174 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 175                               struct fl_flow_mask *mask)
 176 {
 177         const long *lkey = fl_key_get_start(key, mask);
 178         const long *lmask = fl_key_get_start(&mask->key, mask);
 179         long *lmkey = fl_key_get_start(mkey, mask);
 180         int i;
 181
 182         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 183                 *lmkey++ = *lkey++ & *lmask++;
 184 }
 185
 186 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 187                                struct fl_flow_mask *mask)
 188 {
 189         const long *lmask = fl_key_get_start(&mask->key, mask);
 190         const long *ltmplt;
 191         int i;
 192
 193         if (!tmplt)
 194                 return true;
 195         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 196         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 197                 if (~*ltmplt++ & *lmask++)
 198                         return false;
 199         }
 200         return true;
 201 }
 202
 203 static void fl_clear_masked_range(struct fl_flow_key *key,
 204                                   struct fl_flow_mask *mask)
 205 {
 206         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 207 }
 208
 209 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 210                                   struct fl_flow_key *key,
 211                                   struct fl_flow_key *mkey)
 212 {
 213         u16 min_mask, max_mask, min_val, max_val;
 214
 215         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 216         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 217         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 218         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 219
 220         if (min_mask && max_mask) {
 221                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 222                     ntohs(key->tp_range.tp.dst) > max_val)
 223                         return false;
 224
 225                 /* skb does not have min and max values */
 226                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 227                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 228         }
 229         return true;
 230 }
 231
 232 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 233                                   struct fl_flow_key *key,
 234                                   struct fl_flow_key *mkey)
 235 {
 236         u16 min_mask, max_mask, min_val, max_val;
 237
 238         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 239         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 240         min_val = ntohs(filter->key.tp_range.tp_min.src);
 241         max_val = ntohs(filter->key.tp_range.tp_max.src);
 242
 243         if (min_mask && max_mask) {
 244                 if (ntohs(key->tp_range.tp.src) < min_val ||
 245                     ntohs(key->tp_range.tp.src) > max_val)
 246                         return false;
 247
 248                 /* skb does not have min and max values */
 249                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 250                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 251         }
 252         return true;
 253 }
 254
 255 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 256                                          struct fl_flow_key *mkey)
 257 {
 258         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 259                                       mask->filter_ht_params);
 260 }
 261
 262 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 263                                              struct fl_flow_key *mkey,
 264                                              struct fl_flow_key *key)
 265 {
 266         struct cls_fl_filter *filter, *f;
 267
 268         list_for_each_entry_rcu(filter, &mask->filters, list) {
 269                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 270                         continue;
 271
 272                 if (!fl_range_port_src_cmp(filter, key, mkey))
 273                         continue;
 274
 275                 f = __fl_lookup(mask, mkey);
 276                 if (f)
 277                         return f;
 278         }
 279         return NULL;
 280 }
 281
 282 static noinline_for_stack
 283 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 284 {
 285         struct fl_flow_key mkey;
 286
 287         fl_set_masked_key(&mkey, key, mask);
 288         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 289                 return fl_lookup_range(mask, &mkey, key);
 290
 291         return __fl_lookup(mask, &mkey);
 292 }
 293
 294 static u16 fl_ct_info_to_flower_map[] = {
 295         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 296                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 297         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 298                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 299         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 300                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED |
 301                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 302         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 303                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED |
 304                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 305         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 306                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 307 };
 308
 309 TC_INDIRECT_SCOPE int fl_classify(struct sk_buff *skb,
 310                                   const struct tcf_proto *tp,
 311                                   struct tcf_result *res)
 312 {
 313         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 314         bool post_ct = tc_skb_cb(skb)->post_ct;
 315         u16 zone = tc_skb_cb(skb)->zone;
 316         struct fl_flow_key skb_key;
 317         struct fl_flow_mask *mask;
 318         struct cls_fl_filter *f;
 319
 320         list_for_each_entry_rcu(mask, &head->masks, list) {
 321                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 322                 fl_clear_masked_range(&skb_key, mask);
 323
 324                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 325                 /* skb_flow_dissect() does not set n_proto in case an unknown
 326                  * protocol, so do it rather here.
 327                  */
 328                 skb_key.basic.n_proto = skb_protocol(skb, false);
 329                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 330                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 331                                     fl_ct_info_to_flower_map,
 332                                     ARRAY_SIZE(fl_ct_info_to_flower_map),
 333                                     post_ct, zone);
 334                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 335                 skb_flow_dissect(skb, &mask->dissector, &skb_key,
 336                                  FLOW_DISSECTOR_F_STOP_BEFORE_ENCAP);
 337
 338                 f = fl_mask_lookup(mask, &skb_key);
 339                 if (f && !tc_skip_sw(f->flags)) {
 340                         *res = f->res;
 341                         return tcf_exts_exec(skb, &f->exts, res);
 342                 }
 343         }
 344         return -1;
 345 }
 346
 347 static int fl_init(struct tcf_proto *tp)
 348 {
 349         struct cls_fl_head *head;
 350
 351         head = kzalloc(sizeof(*head), GFP_KERNEL);
 352         if (!head)
 353                 return -ENOBUFS;
 354
 355         spin_lock_init(&head->masks_lock);
 356         INIT_LIST_HEAD_RCU(&head->masks);
 357         INIT_LIST_HEAD(&head->hw_filters);
 358         rcu_assign_pointer(tp->root, head);
 359         idr_init(&head->handle_idr);
 360
 361         return rhashtable_init(&head->ht, &mask_ht_params);
 362 }
 363
 364 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 365 {
 366         /* temporary masks don't have their filters list and ht initialized */
 367         if (mask_init_done) {
 368                 WARN_ON(!list_empty(&mask->filters));
 369                 rhashtable_destroy(&mask->ht);
 370         }
 371         kfree(mask);
 372 }
 373
 374 static void fl_mask_free_work(struct work_struct *work)
 375 {
 376         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 377                                                  struct fl_flow_mask, rwork);
 378
 379         fl_mask_free(mask, true);
 380 }
 381
 382 static void fl_uninit_mask_free_work(struct work_struct *work)
 383 {
 384         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 385                                                  struct fl_flow_mask, rwork);
 386
 387         fl_mask_free(mask, false);
 388 }
 389
 390 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 391 {
 392         if (!refcount_dec_and_test(&mask->refcnt))
 393                 return false;
 394
 395         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 396
 397         spin_lock(&head->masks_lock);
 398         list_del_rcu(&mask->list);
 399         spin_unlock(&head->masks_lock);
 400
 401         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 402
 403         return true;
 404 }
 405
 406 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 407 {
 408         /* Flower classifier only changes root pointer during init and destroy.
 409          * Users must obtain reference to tcf_proto instance before calling its
 410          * API, so tp->root pointer is protected from concurrent call to
 411          * fl_destroy() by reference counting.
 412          */
 413         return rcu_dereference_raw(tp->root);
 414 }
 415
 416 static void __fl_destroy_filter(struct cls_fl_filter *f)
 417 {
 418         tcf_exts_destroy(&f->exts);
 419         tcf_exts_put_net(&f->exts);
 420         kfree(f);
 421 }
 422
 423 static void fl_destroy_filter_work(struct work_struct *work)
 424 {
 425         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 426                                         struct cls_fl_filter, rwork);
 427
 428         __fl_destroy_filter(f);
 429 }
 430
 431 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 432                                  bool rtnl_held, struct netlink_ext_ack *extack)
 433 {
 434         struct tcf_block *block = tp->chain->block;
 435         struct flow_cls_offload cls_flower = {};
 436
 437         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 438         cls_flower.command = FLOW_CLS_DESTROY;
 439         cls_flower.cookie = (unsigned long) f;
 440
 441         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 442                             &f->flags, &f->in_hw_count, rtnl_held);
 443
 444 }
 445
 446 static int fl_hw_replace_filter(struct tcf_proto *tp,
 447                                 struct cls_fl_filter *f, bool rtnl_held,
 448                                 struct netlink_ext_ack *extack)
 449 {
 450         struct tcf_block *block = tp->chain->block;
 451         struct flow_cls_offload cls_flower = {};
 452         bool skip_sw = tc_skip_sw(f->flags);
 453         int err = 0;
 454
 455         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 456         if (!cls_flower.rule)
 457                 return -ENOMEM;
 458
 459         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 460         cls_flower.command = FLOW_CLS_REPLACE;
 461         cls_flower.cookie = (unsigned long) f;
 462         cls_flower.rule->match.dissector = &f->mask->dissector;
 463         cls_flower.rule->match.mask = &f->mask->key;
 464         cls_flower.rule->match.key = &f->mkey;
 465         cls_flower.classid = f->res.classid;
 466
 467         err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
 468                                       cls_flower.common.extack);
 469         if (err) {
 470                 kfree(cls_flower.rule);
 471
 472                 return skip_sw ? err : 0;
 473         }
 474
 475         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 476                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 477         tc_cleanup_offload_action(&cls_flower.rule->action);
 478         kfree(cls_flower.rule);
 479
 480         if (err) {
 481                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 482                 return err;
 483         }
 484
 485         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 486                 return -EINVAL;
 487
 488         return 0;
 489 }
 490
 491 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 492                                bool rtnl_held)
 493 {
 494         struct tcf_block *block = tp->chain->block;
 495         struct flow_cls_offload cls_flower = {};
 496
 497         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 498         cls_flower.command = FLOW_CLS_STATS;
 499         cls_flower.cookie = (unsigned long) f;
 500         cls_flower.classid = f->res.classid;
 501
 502         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 503                          rtnl_held);
 504
 505         tcf_exts_hw_stats_update(&f->exts, cls_flower.stats.bytes,
 506                                  cls_flower.stats.pkts,
 507                                  cls_flower.stats.drops,
 508                                  cls_flower.stats.lastused,
 509                                  cls_flower.stats.used_hw_stats,
 510                                  cls_flower.stats.used_hw_stats_valid);
 511 }
 512
 513 static void __fl_put(struct cls_fl_filter *f)
 514 {
 515         if (!refcount_dec_and_test(&f->refcnt))
 516                 return;
 517
 518         if (tcf_exts_get_net(&f->exts))
 519                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 520         else
 521                 __fl_destroy_filter(f);
 522 }
 523
 524 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 525 {
 526         struct cls_fl_filter *f;
 527
 528         rcu_read_lock();
 529         f = idr_find(&head->handle_idr, handle);
 530         if (f && !refcount_inc_not_zero(&f->refcnt))
 531                 f = NULL;
 532         rcu_read_unlock();
 533
 534         return f;
 535 }
 536
 537 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 538                        bool *last, bool rtnl_held,
 539                        struct netlink_ext_ack *extack)
 540 {
 541         struct cls_fl_head *head = fl_head_dereference(tp);
 542
 543         *last = false;
 544
 545         spin_lock(&tp->lock);
 546         if (f->deleted) {
 547                 spin_unlock(&tp->lock);
 548                 return -ENOENT;
 549         }
 550
 551         f->deleted = true;
 552         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 553                                f->mask->filter_ht_params);
 554         idr_remove(&head->handle_idr, f->handle);
 555         list_del_rcu(&f->list);
 556         spin_unlock(&tp->lock);
 557
 558         *last = fl_mask_put(head, f->mask);
 559         if (!tc_skip_hw(f->flags))
 560                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 561         tcf_unbind_filter(tp, &f->res);
 562         __fl_put(f);
 563
 564         return 0;
 565 }
 566
 567 static void fl_destroy_sleepable(struct work_struct *work)
 568 {
 569         struct cls_fl_head *head = container_of(to_rcu_work(work),
 570                                                 struct cls_fl_head,
 571                                                 rwork);
 572
 573         rhashtable_destroy(&head->ht);
 574         kfree(head);
 575         module_put(THIS_MODULE);
 576 }
 577
 578 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 579                        struct netlink_ext_ack *extack)
 580 {
 581         struct cls_fl_head *head = fl_head_dereference(tp);
 582         struct fl_flow_mask *mask, *next_mask;
 583         struct cls_fl_filter *f, *next;
 584         bool last;
 585
 586         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 587                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 588                         __fl_delete(tp, f, &last, rtnl_held, extack);
 589                         if (last)
 590                                 break;
 591                 }
 592         }
 593         idr_destroy(&head->handle_idr);
 594
 595         __module_get(THIS_MODULE);
 596         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 597 }
 598
 599 static void fl_put(struct tcf_proto *tp, void *arg)
 600 {
 601         struct cls_fl_filter *f = arg;
 602
 603         __fl_put(f);
 604 }
 605
 606 static void *fl_get(struct tcf_proto *tp, u32 handle)
 607 {
 608         struct cls_fl_head *head = fl_head_dereference(tp);
 609
 610         return __fl_get(head, handle);
 611 }
 612
 613 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 614         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 615         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 616         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 617                                             .len = IFNAMSIZ },
 618         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 619         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 620         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 621         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 622         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 623         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 624         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 625         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 626         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 627         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 628         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 629         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 630         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 631         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 632         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 633         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 634         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 635         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 636         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 637         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 638         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 639         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 640         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 641         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 642         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 643         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 644         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 645         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 646         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 647         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 648         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 649         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 650         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 660         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 661         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 662         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 663         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 664         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 665         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 666         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 670         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 671         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 672         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 673         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 674         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 675         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 676         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 677         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 678         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 679         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 680         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 681         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 682         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 683         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 684         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 685         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 686         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 687         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 688         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 690         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 692         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 693         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 694         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 695         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 696         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 697         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 698         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 699         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 700         [TCA_FLOWER_KEY_CT_STATE]       =
 701                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 702         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 703                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 704         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 705         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 706         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 707         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 708         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 709                                             .len = 128 / BITS_PER_BYTE },
 710         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 711                                             .len = 128 / BITS_PER_BYTE },
 712         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 713         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 714         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 715         [TCA_FLOWER_KEY_NUM_OF_VLANS]   = { .type = NLA_U8 },
 716         [TCA_FLOWER_KEY_PPPOE_SID]      = { .type = NLA_U16 },
 717         [TCA_FLOWER_KEY_PPP_PROTO]      = { .type = NLA_U16 },
 718         [TCA_FLOWER_KEY_L2TPV3_SID]     = { .type = NLA_U32 },
 719
 720 };
 721
 722 static const struct nla_policy
 723 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 724         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 725                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 726         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 727         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 728         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 729         [TCA_FLOWER_KEY_ENC_OPTS_GTP]           = { .type = NLA_NESTED },
 730 };
 731
 732 static const struct nla_policy
 733 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 734         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 735         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 736         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 737                                                        .len = 128 },
 738 };
 739
 740 static const struct nla_policy
 741 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 742         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 743 };
 744
 745 static const struct nla_policy
 746 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 747         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 748         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 749         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 750         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 751 };
 752
 753 static const struct nla_policy
 754 gtp_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1] = {
 755         [TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]      = { .type = NLA_U8 },
 756         [TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]           = { .type = NLA_U8 },
 757 };
 758
 759 static const struct nla_policy
 760 mpls_stack_entry_policy[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1] = {
 761         [TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]    = { .type = NLA_U8 },
 762         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]      = { .type = NLA_U8 },
 763         [TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]      = { .type = NLA_U8 },
 764         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]       = { .type = NLA_U8 },
 765         [TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]    = { .type = NLA_U32 },
 766 };
 767
 768 static void fl_set_key_val(struct nlattr **tb,
 769                            void *val, int val_type,
 770                            void *mask, int mask_type, int len)
 771 {
 772         if (!tb[val_type])
 773                 return;
 774         nla_memcpy(val, tb[val_type], len);
 775         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 776                 memset(mask, 0xff, len);
 777         else
 778                 nla_memcpy(mask, tb[mask_type], len);
 779 }
 780
 781 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 782                                  struct fl_flow_key *mask,
 783                                  struct netlink_ext_ack *extack)
 784 {
 785         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 786                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 787                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 788         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 789                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 790                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 791         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 792                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 793                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 794         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 795                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 796                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 797
 798         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 799             ntohs(key->tp_range.tp_max.dst) <=
 800             ntohs(key->tp_range.tp_min.dst)) {
 801                 NL_SET_ERR_MSG_ATTR(extack,
 802                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 803                                     "Invalid destination port range (min must be strictly smaller than max)");
 804                 return -EINVAL;
 805         }
 806         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 807             ntohs(key->tp_range.tp_max.src) <=
 808             ntohs(key->tp_range.tp_min.src)) {
 809                 NL_SET_ERR_MSG_ATTR(extack,
 810                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 811                                     "Invalid source port range (min must be strictly smaller than max)");
 812                 return -EINVAL;
 813         }
 814
 815         return 0;
 816 }
 817
 818 static int fl_set_key_mpls_lse(const struct nlattr *nla_lse,
 819                                struct flow_dissector_key_mpls *key_val,
 820                                struct flow_dissector_key_mpls *key_mask,
 821                                struct netlink_ext_ack *extack)
 822 {
 823         struct nlattr *tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1];
 824         struct flow_dissector_mpls_lse *lse_mask;
 825         struct flow_dissector_mpls_lse *lse_val;
 826         u8 lse_index;
 827         u8 depth;
 828         int err;
 829
 830         err = nla_parse_nested(tb, TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX, nla_lse,
 831                                mpls_stack_entry_policy, extack);
 832         if (err < 0)
 833                 return err;
 834
 835         if (!tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]) {
 836                 NL_SET_ERR_MSG(extack, "Missing MPLS option \"depth\"");
 837                 return -EINVAL;
 838         }
 839
 840         depth = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]);
 841
 842         /* LSE depth starts at 1, for consistency with terminology used by
 843          * RFC 3031 (section 3.9), where depth 0 refers to unlabeled packets.
 844          */
 845         if (depth < 1 || depth > FLOW_DIS_MPLS_MAX) {
 846                 NL_SET_ERR_MSG_ATTR(extack,
 847                                     tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH],
 848                                     "Invalid MPLS depth");
 849                 return -EINVAL;
 850         }
 851         lse_index = depth - 1;
 852
 853         dissector_set_mpls_lse(key_val, lse_index);
 854         dissector_set_mpls_lse(key_mask, lse_index);
 855
 856         lse_val = &key_val->ls[lse_index];
 857         lse_mask = &key_mask->ls[lse_index];
 858
 859         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]) {
 860                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]);
 861                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 862         }
 863         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]) {
 864                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]);
 865
 866                 if (bos & ~MPLS_BOS_MASK) {
 867                         NL_SET_ERR_MSG_ATTR(extack,
 868                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS],
 869                                             "Bottom Of Stack (BOS) must be 0 or 1");
 870                         return -EINVAL;
 871                 }
 872                 lse_val->mpls_bos = bos;
 873                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 874         }
 875         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]) {
 876                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]);
 877
 878                 if (tc & ~MPLS_TC_MASK) {
 879                         NL_SET_ERR_MSG_ATTR(extack,
 880                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC],
 881                                             "Traffic Class (TC) must be between 0 and 7");
 882                         return -EINVAL;
 883                 }
 884                 lse_val->mpls_tc = tc;
 885                 lse_mask->mpls_tc = MPLS_TC_MASK;
 886         }
 887         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]) {
 888                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]);
 889
 890                 if (label & ~MPLS_LABEL_MASK) {
 891                         NL_SET_ERR_MSG_ATTR(extack,
 892                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL],
 893                                             "Label must be between 0 and 1048575");
 894                         return -EINVAL;
 895                 }
 896                 lse_val->mpls_label = label;
 897                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 898         }
 899
 900         return 0;
 901 }
 902
 903 static int fl_set_key_mpls_opts(const struct nlattr *nla_mpls_opts,
 904                                 struct flow_dissector_key_mpls *key_val,
 905                                 struct flow_dissector_key_mpls *key_mask,
 906                                 struct netlink_ext_ack *extack)
 907 {
 908         struct nlattr *nla_lse;
 909         int rem;
 910         int err;
 911
 912         if (!(nla_mpls_opts->nla_type & NLA_F_NESTED)) {
 913                 NL_SET_ERR_MSG_ATTR(extack, nla_mpls_opts,
 914                                     "NLA_F_NESTED is missing");
 915                 return -EINVAL;
 916         }
 917
 918         nla_for_each_nested(nla_lse, nla_mpls_opts, rem) {
 919                 if (nla_type(nla_lse) != TCA_FLOWER_KEY_MPLS_OPTS_LSE) {
 920                         NL_SET_ERR_MSG_ATTR(extack, nla_lse,
 921                                             "Invalid MPLS option type");
 922                         return -EINVAL;
 923                 }
 924
 925                 err = fl_set_key_mpls_lse(nla_lse, key_val, key_mask, extack);
 926                 if (err < 0)
 927                         return err;
 928         }
 929         if (rem) {
 930                 NL_SET_ERR_MSG(extack,
 931                                "Bytes leftover after parsing MPLS options");
 932                 return -EINVAL;
 933         }
 934
 935         return 0;
 936 }
 937
 938 static int fl_set_key_mpls(struct nlattr **tb,
 939                            struct flow_dissector_key_mpls *key_val,
 940                            struct flow_dissector_key_mpls *key_mask,
 941                            struct netlink_ext_ack *extack)
 942 {
 943         struct flow_dissector_mpls_lse *lse_mask;
 944         struct flow_dissector_mpls_lse *lse_val;
 945
 946         if (tb[TCA_FLOWER_KEY_MPLS_OPTS]) {
 947                 if (tb[TCA_FLOWER_KEY_MPLS_TTL] ||
 948                     tb[TCA_FLOWER_KEY_MPLS_BOS] ||
 949                     tb[TCA_FLOWER_KEY_MPLS_TC] ||
 950                     tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 951                         NL_SET_ERR_MSG_ATTR(extack,
 952                                             tb[TCA_FLOWER_KEY_MPLS_OPTS],
 953                                             "MPLS label, Traffic Class, Bottom Of Stack and Time To Live must be encapsulated in the MPLS options attribute");
 954                         return -EBADMSG;
 955                 }
 956
 957                 return fl_set_key_mpls_opts(tb[TCA_FLOWER_KEY_MPLS_OPTS],
 958                                             key_val, key_mask, extack);
 959         }
 960
 961         lse_val = &key_val->ls[0];
 962         lse_mask = &key_mask->ls[0];
 963
 964         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 965                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 966                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 967                 dissector_set_mpls_lse(key_val, 0);
 968                 dissector_set_mpls_lse(key_mask, 0);
 969         }
 970         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 971                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 972
 973                 if (bos & ~MPLS_BOS_MASK) {
 974                         NL_SET_ERR_MSG_ATTR(extack,
 975                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
 976                                             "Bottom Of Stack (BOS) must be 0 or 1");
 977                         return -EINVAL;
 978                 }
 979                 lse_val->mpls_bos = bos;
 980                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 981                 dissector_set_mpls_lse(key_val, 0);
 982                 dissector_set_mpls_lse(key_mask, 0);
 983         }
 984         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 985                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 986
 987                 if (tc & ~MPLS_TC_MASK) {
 988                         NL_SET_ERR_MSG_ATTR(extack,
 989                                             tb[TCA_FLOWER_KEY_MPLS_TC],
 990                                             "Traffic Class (TC) must be between 0 and 7");
 991                         return -EINVAL;
 992                 }
 993                 lse_val->mpls_tc = tc;
 994                 lse_mask->mpls_tc = MPLS_TC_MASK;
 995                 dissector_set_mpls_lse(key_val, 0);
 996                 dissector_set_mpls_lse(key_mask, 0);
 997         }
 998         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 999                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
1000
1001                 if (label & ~MPLS_LABEL_MASK) {
1002                         NL_SET_ERR_MSG_ATTR(extack,
1003                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
1004                                             "Label must be between 0 and 1048575");
1005                         return -EINVAL;
1006                 }
1007                 lse_val->mpls_label = label;
1008                 lse_mask->mpls_label = MPLS_LABEL_MASK;
1009                 dissector_set_mpls_lse(key_val, 0);
1010                 dissector_set_mpls_lse(key_mask, 0);
1011         }
1012         return 0;
1013 }
1014
1015 static void fl_set_key_vlan(struct nlattr **tb,
1016                             __be16 ethertype,
1017                             int vlan_id_key, int vlan_prio_key,
1018                             int vlan_next_eth_type_key,
1019                             struct flow_dissector_key_vlan *key_val,
1020                             struct flow_dissector_key_vlan *key_mask)
1021 {
1022 #define VLAN_PRIORITY_MASK      0x7
1023
1024         if (tb[vlan_id_key]) {
1025                 key_val->vlan_id =
1026                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1027                 key_mask->vlan_id = VLAN_VID_MASK;
1028         }
1029         if (tb[vlan_prio_key]) {
1030                 key_val->vlan_priority =
1031                         nla_get_u8(tb[vlan_prio_key]) &
1032                         VLAN_PRIORITY_MASK;
1033                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1034         }
1035         if (ethertype) {
1036                 key_val->vlan_tpid = ethertype;
1037                 key_mask->vlan_tpid = cpu_to_be16(~0);
1038         }
1039         if (tb[vlan_next_eth_type_key]) {
1040                 key_val->vlan_eth_type =
1041                         nla_get_be16(tb[vlan_next_eth_type_key]);
1042                 key_mask->vlan_eth_type = cpu_to_be16(~0);
1043         }
1044 }
1045
1046 static void fl_set_key_pppoe(struct nlattr **tb,
1047                              struct flow_dissector_key_pppoe *key_val,
1048                              struct flow_dissector_key_pppoe *key_mask,
1049                              struct fl_flow_key *key,
1050                              struct fl_flow_key *mask)
1051 {
1052         /* key_val::type must be set to ETH_P_PPP_SES
1053          * because ETH_P_PPP_SES was stored in basic.n_proto
1054          * which might get overwritten by ppp_proto
1055          * or might be set to 0, the role of key_val::type
1056          * is simmilar to vlan_key::tpid
1057          */
1058         key_val->type = htons(ETH_P_PPP_SES);
1059         key_mask->type = cpu_to_be16(~0);
1060
1061         if (tb[TCA_FLOWER_KEY_PPPOE_SID]) {
1062                 key_val->session_id =
1063                         nla_get_be16(tb[TCA_FLOWER_KEY_PPPOE_SID]);
1064                 key_mask->session_id = cpu_to_be16(~0);
1065         }
1066         if (tb[TCA_FLOWER_KEY_PPP_PROTO]) {
1067                 key_val->ppp_proto =
1068                         nla_get_be16(tb[TCA_FLOWER_KEY_PPP_PROTO]);
1069                 key_mask->ppp_proto = cpu_to_be16(~0);
1070
1071                 if (key_val->ppp_proto == htons(PPP_IP)) {
1072                         key->basic.n_proto = htons(ETH_P_IP);
1073                         mask->basic.n_proto = cpu_to_be16(~0);
1074                 } else if (key_val->ppp_proto == htons(PPP_IPV6)) {
1075                         key->basic.n_proto = htons(ETH_P_IPV6);
1076                         mask->basic.n_proto = cpu_to_be16(~0);
1077                 } else if (key_val->ppp_proto == htons(PPP_MPLS_UC)) {
1078                         key->basic.n_proto = htons(ETH_P_MPLS_UC);
1079                         mask->basic.n_proto = cpu_to_be16(~0);
1080                 } else if (key_val->ppp_proto == htons(PPP_MPLS_MC)) {
1081                         key->basic.n_proto = htons(ETH_P_MPLS_MC);
1082                         mask->basic.n_proto = cpu_to_be16(~0);
1083                 }
1084         } else {
1085                 key->basic.n_proto = 0;
1086                 mask->basic.n_proto = cpu_to_be16(0);
1087         }
1088 }
1089
1090 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1091                             u32 *dissector_key, u32 *dissector_mask,
1092                             u32 flower_flag_bit, u32 dissector_flag_bit)
1093 {
1094         if (flower_mask & flower_flag_bit) {
1095                 *dissector_mask |= dissector_flag_bit;
1096                 if (flower_key & flower_flag_bit)
1097                         *dissector_key |= dissector_flag_bit;
1098         }
1099 }
1100
1101 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1102                             u32 *flags_mask, struct netlink_ext_ack *extack)
1103 {
1104         u32 key, mask;
1105
1106         /* mask is mandatory for flags */
1107         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1108                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1109                 return -EINVAL;
1110         }
1111
1112         key = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS]));
1113         mask = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1114
1115         *flags_key  = 0;
1116         *flags_mask = 0;
1117
1118         fl_set_key_flag(key, mask, flags_key, flags_mask,
1119                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1120         fl_set_key_flag(key, mask, flags_key, flags_mask,
1121                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1122                         FLOW_DIS_FIRST_FRAG);
1123
1124         return 0;
1125 }
1126
1127 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1128                           struct flow_dissector_key_ip *key,
1129                           struct flow_dissector_key_ip *mask)
1130 {
1131         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1132         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1133         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1134         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1135
1136         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1137         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1138 }
1139
1140 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1141                              int depth, int option_len,
1142                              struct netlink_ext_ack *extack)
1143 {
1144         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1145         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1146         struct geneve_opt *opt;
1147         int err, data_len = 0;
1148
1149         if (option_len > sizeof(struct geneve_opt))
1150                 data_len = option_len - sizeof(struct geneve_opt);
1151
1152         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1153         memset(opt, 0xff, option_len);
1154         opt->length = data_len / 4;
1155         opt->r1 = 0;
1156         opt->r2 = 0;
1157         opt->r3 = 0;
1158
1159         /* If no mask has been prodived we assume an exact match. */
1160         if (!depth)
1161                 return sizeof(struct geneve_opt) + data_len;
1162
1163         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1164                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1165                 return -EINVAL;
1166         }
1167
1168         err = nla_parse_nested_deprecated(tb,
1169                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1170                                           nla, geneve_opt_policy, extack);
1171         if (err < 0)
1172                 return err;
1173
1174         /* We are not allowed to omit any of CLASS, TYPE or DATA
1175          * fields from the key.
1176          */
1177         if (!option_len &&
1178             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1179              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1180              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1181                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1182                 return -EINVAL;
1183         }
1184
1185         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1186          * for the mask.
1187          */
1188         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1189                 int new_len = key->enc_opts.len;
1190
1191                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1192                 data_len = nla_len(data);
1193                 if (data_len < 4) {
1194                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1195                         return -ERANGE;
1196                 }
1197                 if (data_len % 4) {
1198                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1199                         return -ERANGE;
1200                 }
1201
1202                 new_len += sizeof(struct geneve_opt) + data_len;
1203                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1204                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1205                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1206                         return -ERANGE;
1207                 }
1208                 opt->length = data_len / 4;
1209                 memcpy(opt->opt_data, nla_data(data), data_len);
1210         }
1211
1212         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1213                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1214                 opt->opt_class = nla_get_be16(class);
1215         }
1216
1217         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1218                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1219                 opt->type = nla_get_u8(type);
1220         }
1221
1222         return sizeof(struct geneve_opt) + data_len;
1223 }
1224
1225 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1226                             int depth, int option_len,
1227                             struct netlink_ext_ack *extack)
1228 {
1229         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1230         struct vxlan_metadata *md;
1231         int err;
1232
1233         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1234         memset(md, 0xff, sizeof(*md));
1235
1236         if (!depth)
1237                 return sizeof(*md);
1238
1239         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1240                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1241                 return -EINVAL;
1242         }
1243
1244         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1245                                vxlan_opt_policy, extack);
1246         if (err < 0)
1247                 return err;
1248
1249         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1250                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1251                 return -EINVAL;
1252         }
1253
1254         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1255                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1256                 md->gbp &= VXLAN_GBP_MASK;
1257         }
1258
1259         return sizeof(*md);
1260 }
1261
1262 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1263                              int depth, int option_len,
1264                              struct netlink_ext_ack *extack)
1265 {
1266         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1267         struct erspan_metadata *md;
1268         int err;
1269
1270         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1271         memset(md, 0xff, sizeof(*md));
1272         md->version = 1;
1273
1274         if (!depth)
1275                 return sizeof(*md);
1276
1277         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1278                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1279                 return -EINVAL;
1280         }
1281
1282         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1283                                erspan_opt_policy, extack);
1284         if (err < 0)
1285                 return err;
1286
1287         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1288                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1289                 return -EINVAL;
1290         }
1291
1292         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1293                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1294
1295         if (md->version == 1) {
1296                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1297                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1298                         return -EINVAL;
1299                 }
1300                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1301                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1302                         memset(&md->u, 0x00, sizeof(md->u));
1303                         md->u.index = nla_get_be32(nla);
1304                 }
1305         } else if (md->version == 2) {
1306                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1307                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1308                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1309                         return -EINVAL;
1310                 }
1311                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1312                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1313                         md->u.md2.dir = nla_get_u8(nla);
1314                 }
1315                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1316                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1317                         set_hwid(&md->u.md2, nla_get_u8(nla));
1318                 }
1319         } else {
1320                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1321                 return -EINVAL;
1322         }
1323
1324         return sizeof(*md);
1325 }
1326
1327 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1328                           int depth, int option_len,
1329                           struct netlink_ext_ack *extack)
1330 {
1331         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1332         struct gtp_pdu_session_info *sinfo;
1333         u8 len = key->enc_opts.len;
1334         int err;
1335
1336         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1337         memset(sinfo, 0xff, option_len);
1338
1339         if (!depth)
1340                 return sizeof(*sinfo);
1341
1342         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1343                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1344                 return -EINVAL;
1345         }
1346
1347         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1348                                gtp_opt_policy, extack);
1349         if (err < 0)
1350                 return err;
1351
1352         if (!option_len &&
1353             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1354              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1355                 NL_SET_ERR_MSG_MOD(extack,
1356                                    "Missing tunnel key gtp option pdu type or qfi");
1357                 return -EINVAL;
1358         }
1359
1360         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1361                 sinfo->pdu_type =
1362                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1363
1364         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1365                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1366
1367         return sizeof(*sinfo);
1368 }
1369
1370 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1371                           struct fl_flow_key *mask,
1372                           struct netlink_ext_ack *extack)
1373 {
1374         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1375         int err, option_len, key_depth, msk_depth = 0;
1376
1377         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1378                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1379                                              enc_opts_policy, extack);
1380         if (err)
1381                 return err;
1382
1383         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1384
1385         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1386                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1387                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1388                                                      enc_opts_policy, extack);
1389                 if (err)
1390                         return err;
1391
1392                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1393                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1394                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1395                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1396                         return -EINVAL;
1397                 }
1398         }
1399
1400         nla_for_each_attr(nla_opt_key, nla_enc_key,
1401                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1402                 switch (nla_type(nla_opt_key)) {
1403                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1404                         if (key->enc_opts.dst_opt_type &&
1405                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1406                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1407                                 return -EINVAL;
1408                         }
1409                         option_len = 0;
1410                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1411                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1412                                                        key_depth, option_len,
1413                                                        extack);
1414                         if (option_len < 0)
1415                                 return option_len;
1416
1417                         key->enc_opts.len += option_len;
1418                         /* At the same time we need to parse through the mask
1419                          * in order to verify exact and mask attribute lengths.
1420                          */
1421                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1422                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1423                                                        msk_depth, option_len,
1424                                                        extack);
1425                         if (option_len < 0)
1426                                 return option_len;
1427
1428                         mask->enc_opts.len += option_len;
1429                         if (key->enc_opts.len != mask->enc_opts.len) {
1430                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1431                                 return -EINVAL;
1432                         }
1433                         break;
1434                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1435                         if (key->enc_opts.dst_opt_type) {
1436                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1437                                 return -EINVAL;
1438                         }
1439                         option_len = 0;
1440                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1441                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1442                                                       key_depth, option_len,
1443                                                       extack);
1444                         if (option_len < 0)
1445                                 return option_len;
1446
1447                         key->enc_opts.len += option_len;
1448                         /* At the same time we need to parse through the mask
1449                          * in order to verify exact and mask attribute lengths.
1450                          */
1451                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1452                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1453                                                       msk_depth, option_len,
1454                                                       extack);
1455                         if (option_len < 0)
1456                                 return option_len;
1457
1458                         mask->enc_opts.len += option_len;
1459                         if (key->enc_opts.len != mask->enc_opts.len) {
1460                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1461                                 return -EINVAL;
1462                         }
1463                         break;
1464                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1465                         if (key->enc_opts.dst_opt_type) {
1466                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1467                                 return -EINVAL;
1468                         }
1469                         option_len = 0;
1470                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1471                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1472                                                        key_depth, option_len,
1473                                                        extack);
1474                         if (option_len < 0)
1475                                 return option_len;
1476
1477                         key->enc_opts.len += option_len;
1478                         /* At the same time we need to parse through the mask
1479                          * in order to verify exact and mask attribute lengths.
1480                          */
1481                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1482                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1483                                                        msk_depth, option_len,
1484                                                        extack);
1485                         if (option_len < 0)
1486                                 return option_len;
1487
1488                         mask->enc_opts.len += option_len;
1489                         if (key->enc_opts.len != mask->enc_opts.len) {
1490                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1491                                 return -EINVAL;
1492                         }
1493                         break;
1494                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1495                         if (key->enc_opts.dst_opt_type) {
1496                                 NL_SET_ERR_MSG_MOD(extack,
1497                                                    "Duplicate type for gtp options");
1498                                 return -EINVAL;
1499                         }
1500                         option_len = 0;
1501                         key->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1502                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1503                                                     key_depth, option_len,
1504                                                     extack);
1505                         if (option_len < 0)
1506                                 return option_len;
1507
1508                         key->enc_opts.len += option_len;
1509                         /* At the same time we need to parse through the mask
1510                          * in order to verify exact and mask attribute lengths.
1511                          */
1512                         mask->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1513                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1514                                                     msk_depth, option_len,
1515                                                     extack);
1516                         if (option_len < 0)
1517                                 return option_len;
1518
1519                         mask->enc_opts.len += option_len;
1520                         if (key->enc_opts.len != mask->enc_opts.len) {
1521                                 NL_SET_ERR_MSG_MOD(extack,
1522                                                    "Key and mask miss aligned");
1523                                 return -EINVAL;
1524                         }
1525                         break;
1526                 default:
1527                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1528                         return -EINVAL;
1529                 }
1530
1531                 if (!msk_depth)
1532                         continue;
1533
1534                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1535                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1536                         return -EINVAL;
1537                 }
1538                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1539         }
1540
1541         return 0;
1542 }
1543
1544 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1545                                 struct netlink_ext_ack *extack)
1546 {
1547         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1548                 NL_SET_ERR_MSG_ATTR(extack, tb,
1549                                     "no trk, so no other flag can be set");
1550                 return -EINVAL;
1551         }
1552
1553         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1554             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1555                 NL_SET_ERR_MSG_ATTR(extack, tb,
1556                                     "new and est are mutually exclusive");
1557                 return -EINVAL;
1558         }
1559
1560         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1561             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1562                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1563                 NL_SET_ERR_MSG_ATTR(extack, tb,
1564                                     "when inv is set, only trk may be set");
1565                 return -EINVAL;
1566         }
1567
1568         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1569             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1570                 NL_SET_ERR_MSG_ATTR(extack, tb,
1571                                     "new and rpl are mutually exclusive");
1572                 return -EINVAL;
1573         }
1574
1575         return 0;
1576 }
1577
1578 static int fl_set_key_ct(struct nlattr **tb,
1579                          struct flow_dissector_key_ct *key,
1580                          struct flow_dissector_key_ct *mask,
1581                          struct netlink_ext_ack *extack)
1582 {
1583         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1584                 int err;
1585
1586                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1587                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1588                         return -EOPNOTSUPP;
1589                 }
1590                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1591                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1592                                sizeof(key->ct_state));
1593
1594                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1595                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1596                                            extack);
1597                 if (err)
1598                         return err;
1599
1600         }
1601         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1602                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1603                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1604                         return -EOPNOTSUPP;
1605                 }
1606                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1607                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1608                                sizeof(key->ct_zone));
1609         }
1610         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1611                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1612                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1613                         return -EOPNOTSUPP;
1614                 }
1615                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1616                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1617                                sizeof(key->ct_mark));
1618         }
1619         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1620                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1621                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1622                         return -EOPNOTSUPP;
1623                 }
1624                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1625                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1626                                sizeof(key->ct_labels));
1627         }
1628
1629         return 0;
1630 }
1631
1632 static bool is_vlan_key(struct nlattr *tb, __be16 *ethertype,
1633                         struct fl_flow_key *key, struct fl_flow_key *mask,
1634                         int vthresh)
1635 {
1636         const bool good_num_of_vlans = key->num_of_vlans.num_of_vlans > vthresh;
1637
1638         if (!tb) {
1639                 *ethertype = 0;
1640                 return good_num_of_vlans;
1641         }
1642
1643         *ethertype = nla_get_be16(tb);
1644         if (good_num_of_vlans || eth_type_vlan(*ethertype))
1645                 return true;
1646
1647         key->basic.n_proto = *ethertype;
1648         mask->basic.n_proto = cpu_to_be16(~0);
1649         return false;
1650 }
1651
1652 static int fl_set_key(struct net *net, struct nlattr **tb,
1653                       struct fl_flow_key *key, struct fl_flow_key *mask,
1654                       struct netlink_ext_ack *extack)
1655 {
1656         __be16 ethertype;
1657         int ret = 0;
1658
1659         if (tb[TCA_FLOWER_INDEV]) {
1660                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1661                 if (err < 0)
1662                         return err;
1663                 key->meta.ingress_ifindex = err;
1664                 mask->meta.ingress_ifindex = 0xffffffff;
1665         }
1666
1667         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1668                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1669                        sizeof(key->eth.dst));
1670         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1671                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1672                        sizeof(key->eth.src));
1673         fl_set_key_val(tb, &key->num_of_vlans,
1674                        TCA_FLOWER_KEY_NUM_OF_VLANS,
1675                        &mask->num_of_vlans,
1676                        TCA_FLOWER_UNSPEC,
1677                        sizeof(key->num_of_vlans));
1678
1679         if (is_vlan_key(tb[TCA_FLOWER_KEY_ETH_TYPE], &ethertype, key, mask, 0)) {
1680                 fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1681                                 TCA_FLOWER_KEY_VLAN_PRIO,
1682                                 TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1683                                 &key->vlan, &mask->vlan);
1684
1685                 if (is_vlan_key(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE],
1686                                 &ethertype, key, mask, 1)) {
1687                         fl_set_key_vlan(tb, ethertype,
1688                                         TCA_FLOWER_KEY_CVLAN_ID,
1689                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1690                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1691                                         &key->cvlan, &mask->cvlan);
1692                         fl_set_key_val(tb, &key->basic.n_proto,
1693                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1694                                        &mask->basic.n_proto,
1695                                        TCA_FLOWER_UNSPEC,
1696                                        sizeof(key->basic.n_proto));
1697                 }
1698         }
1699
1700         if (key->basic.n_proto == htons(ETH_P_PPP_SES))
1701                 fl_set_key_pppoe(tb, &key->pppoe, &mask->pppoe, key, mask);
1702
1703         if (key->basic.n_proto == htons(ETH_P_IP) ||
1704             key->basic.n_proto == htons(ETH_P_IPV6)) {
1705                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1706                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1707                                sizeof(key->basic.ip_proto));
1708                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1709         }
1710
1711         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1712                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1713                 mask->control.addr_type = ~0;
1714                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1715                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1716                                sizeof(key->ipv4.src));
1717                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1718                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1719                                sizeof(key->ipv4.dst));
1720         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1721                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1722                 mask->control.addr_type = ~0;
1723                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1724                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1725                                sizeof(key->ipv6.src));
1726                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1727                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1728                                sizeof(key->ipv6.dst));
1729         }
1730
1731         if (key->basic.ip_proto == IPPROTO_TCP) {
1732                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1733                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1734                                sizeof(key->tp.src));
1735                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1736                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1737                                sizeof(key->tp.dst));
1738                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1739                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1740                                sizeof(key->tcp.flags));
1741         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1742                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1743                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1744                                sizeof(key->tp.src));
1745                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1746                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1747                                sizeof(key->tp.dst));
1748         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1749                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1750                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1751                                sizeof(key->tp.src));
1752                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1753                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1754                                sizeof(key->tp.dst));
1755         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1756                    key->basic.ip_proto == IPPROTO_ICMP) {
1757                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1758                                &mask->icmp.type,
1759                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1760                                sizeof(key->icmp.type));
1761                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1762                                &mask->icmp.code,
1763                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1764                                sizeof(key->icmp.code));
1765         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1766                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1767                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1768                                &mask->icmp.type,
1769                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1770                                sizeof(key->icmp.type));
1771                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1772                                &mask->icmp.code,
1773                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1774                                sizeof(key->icmp.code));
1775         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1776                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1777                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1778                 if (ret)
1779                         return ret;
1780         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1781                    key->basic.n_proto == htons(ETH_P_RARP)) {
1782                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1783                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1784                                sizeof(key->arp.sip));
1785                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1786                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1787                                sizeof(key->arp.tip));
1788                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1789                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1790                                sizeof(key->arp.op));
1791                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1792                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1793                                sizeof(key->arp.sha));
1794                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1795                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1796                                sizeof(key->arp.tha));
1797         } else if (key->basic.ip_proto == IPPROTO_L2TP) {
1798                 fl_set_key_val(tb, &key->l2tpv3.session_id,
1799                                TCA_FLOWER_KEY_L2TPV3_SID,
1800                                &mask->l2tpv3.session_id, TCA_FLOWER_UNSPEC,
1801                                sizeof(key->l2tpv3.session_id));
1802         }
1803
1804         if (key->basic.ip_proto == IPPROTO_TCP ||
1805             key->basic.ip_proto == IPPROTO_UDP ||
1806             key->basic.ip_proto == IPPROTO_SCTP) {
1807                 ret = fl_set_key_port_range(tb, key, mask, extack);
1808                 if (ret)
1809                         return ret;
1810         }
1811
1812         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1813             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1814                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1815                 mask->enc_control.addr_type = ~0;
1816                 fl_set_key_val(tb, &key->enc_ipv4.src,
1817                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1818                                &mask->enc_ipv4.src,
1819                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1820                                sizeof(key->enc_ipv4.src));
1821                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1822                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1823                                &mask->enc_ipv4.dst,
1824                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1825                                sizeof(key->enc_ipv4.dst));
1826         }
1827
1828         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1829             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1830                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1831                 mask->enc_control.addr_type = ~0;
1832                 fl_set_key_val(tb, &key->enc_ipv6.src,
1833                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1834                                &mask->enc_ipv6.src,
1835                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1836                                sizeof(key->enc_ipv6.src));
1837                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1838                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1839                                &mask->enc_ipv6.dst,
1840                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1841                                sizeof(key->enc_ipv6.dst));
1842         }
1843
1844         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1845                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1846                        sizeof(key->enc_key_id.keyid));
1847
1848         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1849                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1850                        sizeof(key->enc_tp.src));
1851
1852         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1853                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1854                        sizeof(key->enc_tp.dst));
1855
1856         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1857
1858         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1859                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1860                        sizeof(key->hash.hash));
1861
1862         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1863                 ret = fl_set_enc_opt(tb, key, mask, extack);
1864                 if (ret)
1865                         return ret;
1866         }
1867
1868         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1869         if (ret)
1870                 return ret;
1871
1872         if (tb[TCA_FLOWER_KEY_FLAGS])
1873                 ret = fl_set_key_flags(tb, &key->control.flags,
1874                                        &mask->control.flags, extack);
1875
1876         return ret;
1877 }
1878
1879 static void fl_mask_copy(struct fl_flow_mask *dst,
1880                          struct fl_flow_mask *src)
1881 {
1882         const void *psrc = fl_key_get_start(&src->key, src);
1883         void *pdst = fl_key_get_start(&dst->key, src);
1884
1885         memcpy(pdst, psrc, fl_mask_range(src));
1886         dst->range = src->range;
1887 }
1888
1889 static const struct rhashtable_params fl_ht_params = {
1890         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1891         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1892         .automatic_shrinking = true,
1893 };
1894
1895 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1896 {
1897         mask->filter_ht_params = fl_ht_params;
1898         mask->filter_ht_params.key_len = fl_mask_range(mask);
1899         mask->filter_ht_params.key_offset += mask->range.start;
1900
1901         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1902 }
1903
1904 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1905 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1906
1907 #define FL_KEY_IS_MASKED(mask, member)                                          \
1908         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1909                    0, FL_KEY_MEMBER_SIZE(member))                               \
1910
1911 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1912         do {                                                                    \
1913                 keys[cnt].key_id = id;                                          \
1914                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1915                 cnt++;                                                          \
1916         } while(0);
1917
1918 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1919         do {                                                                    \
1920                 if (FL_KEY_IS_MASKED(mask, member))                             \
1921                         FL_KEY_SET(keys, cnt, id, member);                      \
1922         } while(0);
1923
1924 static void fl_init_dissector(struct flow_dissector *dissector,
1925                               struct fl_flow_key *mask)
1926 {
1927         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1928         size_t cnt = 0;
1929
1930         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1931                              FLOW_DISSECTOR_KEY_META, meta);
1932         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1933         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1934         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1935                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1936         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1937                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1938         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1939                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1940         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1941                              FLOW_DISSECTOR_KEY_PORTS, tp);
1942         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1943                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1944         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1945                              FLOW_DISSECTOR_KEY_IP, ip);
1946         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1947                              FLOW_DISSECTOR_KEY_TCP, tcp);
1948         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1949                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1950         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1951                              FLOW_DISSECTOR_KEY_ARP, arp);
1952         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1953                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1954         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1955                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1956         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1957                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1958         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1959                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1960         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1961                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1962         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1963                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1964         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1965             FL_KEY_IS_MASKED(mask, enc_ipv6))
1966                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1967                            enc_control);
1968         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1969                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1970         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1971                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1972         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1973                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1974         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1975                              FLOW_DISSECTOR_KEY_CT, ct);
1976         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1977                              FLOW_DISSECTOR_KEY_HASH, hash);
1978         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1979                              FLOW_DISSECTOR_KEY_NUM_OF_VLANS, num_of_vlans);
1980         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1981                              FLOW_DISSECTOR_KEY_PPPOE, pppoe);
1982         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1983                              FLOW_DISSECTOR_KEY_L2TPV3, l2tpv3);
1984
1985         skb_flow_dissector_init(dissector, keys, cnt);
1986 }
1987
1988 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1989                                                struct fl_flow_mask *mask)
1990 {
1991         struct fl_flow_mask *newmask;
1992         int err;
1993
1994         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1995         if (!newmask)
1996                 return ERR_PTR(-ENOMEM);
1997
1998         fl_mask_copy(newmask, mask);
1999
2000         if ((newmask->key.tp_range.tp_min.dst &&
2001              newmask->key.tp_range.tp_max.dst) ||
2002             (newmask->key.tp_range.tp_min.src &&
2003              newmask->key.tp_range.tp_max.src))
2004                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
2005
2006         err = fl_init_mask_hashtable(newmask);
2007         if (err)
2008                 goto errout_free;
2009
2010         fl_init_dissector(&newmask->dissector, &newmask->key);
2011
2012         INIT_LIST_HEAD_RCU(&newmask->filters);
2013
2014         refcount_set(&newmask->refcnt, 1);
2015         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
2016                                       &newmask->ht_node, mask_ht_params);
2017         if (err)
2018                 goto errout_destroy;
2019
2020         spin_lock(&head->masks_lock);
2021         list_add_tail_rcu(&newmask->list, &head->masks);
2022         spin_unlock(&head->masks_lock);
2023
2024         return newmask;
2025
2026 errout_destroy:
2027         rhashtable_destroy(&newmask->ht);
2028 errout_free:
2029         kfree(newmask);
2030
2031         return ERR_PTR(err);
2032 }
2033
2034 static int fl_check_assign_mask(struct cls_fl_head *head,
2035                                 struct cls_fl_filter *fnew,
2036                                 struct cls_fl_filter *fold,
2037                                 struct fl_flow_mask *mask)
2038 {
2039         struct fl_flow_mask *newmask;
2040         int ret = 0;
2041
2042         rcu_read_lock();
2043
2044         /* Insert mask as temporary node to prevent concurrent creation of mask
2045          * with same key. Any concurrent lookups with same key will return
2046          * -EAGAIN because mask's refcnt is zero.
2047          */
2048         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
2049                                                        &mask->ht_node,
2050                                                        mask_ht_params);
2051         if (!fnew->mask) {
2052                 rcu_read_unlock();
2053
2054                 if (fold) {
2055                         ret = -EINVAL;
2056                         goto errout_cleanup;
2057                 }
2058
2059                 newmask = fl_create_new_mask(head, mask);
2060                 if (IS_ERR(newmask)) {
2061                         ret = PTR_ERR(newmask);
2062                         goto errout_cleanup;
2063                 }
2064
2065                 fnew->mask = newmask;
2066                 return 0;
2067         } else if (IS_ERR(fnew->mask)) {
2068                 ret = PTR_ERR(fnew->mask);
2069         } else if (fold && fold->mask != fnew->mask) {
2070                 ret = -EINVAL;
2071         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
2072                 /* Mask was deleted concurrently, try again */
2073                 ret = -EAGAIN;
2074         }
2075         rcu_read_unlock();
2076         return ret;
2077
2078 errout_cleanup:
2079         rhashtable_remove_fast(&head->ht, &mask->ht_node,
2080                                mask_ht_params);
2081         return ret;
2082 }
2083
2084 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
2085                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
2086                         unsigned long base, struct nlattr **tb,
2087                         struct nlattr *est,
2088                         struct fl_flow_tmplt *tmplt,
2089                         u32 flags, u32 fl_flags,
2090                         struct netlink_ext_ack *extack)
2091 {
2092         int err;
2093
2094         err = tcf_exts_validate_ex(net, tp, tb, est, &f->exts, flags,
2095                                    fl_flags, extack);
2096         if (err < 0)
2097                 return err;
2098
2099         if (tb[TCA_FLOWER_CLASSID]) {
2100                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2101                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2102                         rtnl_lock();
2103                 tcf_bind_filter(tp, &f->res, base);
2104                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2105                         rtnl_unlock();
2106         }
2107
2108         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
2109         if (err)
2110                 return err;
2111
2112         fl_mask_update_range(mask);
2113         fl_set_masked_key(&f->mkey, &f->key, mask);
2114
2115         if (!fl_mask_fits_tmplt(tmplt, mask)) {
2116                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2117                 return -EINVAL;
2118         }
2119
2120         return 0;
2121 }
2122
2123 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2124                                struct cls_fl_filter *fold,
2125                                bool *in_ht)
2126 {
2127         struct fl_flow_mask *mask = fnew->mask;
2128         int err;
2129
2130         err = rhashtable_lookup_insert_fast(&mask->ht,
2131                                             &fnew->ht_node,
2132                                             mask->filter_ht_params);
2133         if (err) {
2134                 *in_ht = false;
2135                 /* It is okay if filter with same key exists when
2136                  * overwriting.
2137                  */
2138                 return fold && err == -EEXIST ? 0 : err;
2139         }
2140
2141         *in_ht = true;
2142         return 0;
2143 }
2144
2145 static int fl_change(struct net *net, struct sk_buff *in_skb,
2146                      struct tcf_proto *tp, unsigned long base,
2147                      u32 handle, struct nlattr **tca,
2148                      void **arg, u32 flags,
2149                      struct netlink_ext_ack *extack)
2150 {
2151         struct cls_fl_head *head = fl_head_dereference(tp);
2152         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2153         struct cls_fl_filter *fold = *arg;
2154         struct cls_fl_filter *fnew;
2155         struct fl_flow_mask *mask;
2156         struct nlattr **tb;
2157         bool in_ht;
2158         int err;
2159
2160         if (!tca[TCA_OPTIONS]) {
2161                 err = -EINVAL;
2162                 goto errout_fold;
2163         }
2164
2165         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2166         if (!mask) {
2167                 err = -ENOBUFS;
2168                 goto errout_fold;
2169         }
2170
2171         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2172         if (!tb) {
2173                 err = -ENOBUFS;
2174                 goto errout_mask_alloc;
2175         }
2176
2177         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2178                                           tca[TCA_OPTIONS], fl_policy, NULL);
2179         if (err < 0)
2180                 goto errout_tb;
2181
2182         if (fold && handle && fold->handle != handle) {
2183                 err = -EINVAL;
2184                 goto errout_tb;
2185         }
2186
2187         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2188         if (!fnew) {
2189                 err = -ENOBUFS;
2190                 goto errout_tb;
2191         }
2192         INIT_LIST_HEAD(&fnew->hw_list);
2193         refcount_set(&fnew->refcnt, 1);
2194
2195         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
2196         if (err < 0)
2197                 goto errout;
2198
2199         if (tb[TCA_FLOWER_FLAGS]) {
2200                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2201
2202                 if (!tc_flags_valid(fnew->flags)) {
2203                         err = -EINVAL;
2204                         goto errout;
2205                 }
2206         }
2207
2208         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2209                            tp->chain->tmplt_priv, flags, fnew->flags,
2210                            extack);
2211         if (err)
2212                 goto errout;
2213
2214         err = fl_check_assign_mask(head, fnew, fold, mask);
2215         if (err)
2216                 goto errout;
2217
2218         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2219         if (err)
2220                 goto errout_mask;
2221
2222         if (!tc_skip_hw(fnew->flags)) {
2223                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2224                 if (err)
2225                         goto errout_ht;
2226         }
2227
2228         if (!tc_in_hw(fnew->flags))
2229                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2230
2231         spin_lock(&tp->lock);
2232
2233         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2234          * proto again or create new one, if necessary.
2235          */
2236         if (tp->deleting) {
2237                 err = -EAGAIN;
2238                 goto errout_hw;
2239         }
2240
2241         if (fold) {
2242                 /* Fold filter was deleted concurrently. Retry lookup. */
2243                 if (fold->deleted) {
2244                         err = -EAGAIN;
2245                         goto errout_hw;
2246                 }
2247
2248                 fnew->handle = handle;
2249
2250                 if (!in_ht) {
2251                         struct rhashtable_params params =
2252                                 fnew->mask->filter_ht_params;
2253
2254                         err = rhashtable_insert_fast(&fnew->mask->ht,
2255                                                      &fnew->ht_node,
2256                                                      params);
2257                         if (err)
2258                                 goto errout_hw;
2259                         in_ht = true;
2260                 }
2261
2262                 refcount_inc(&fnew->refcnt);
2263                 rhashtable_remove_fast(&fold->mask->ht,
2264                                        &fold->ht_node,
2265                                        fold->mask->filter_ht_params);
2266                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2267                 list_replace_rcu(&fold->list, &fnew->list);
2268                 fold->deleted = true;
2269
2270                 spin_unlock(&tp->lock);
2271
2272                 fl_mask_put(head, fold->mask);
2273                 if (!tc_skip_hw(fold->flags))
2274                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2275                 tcf_unbind_filter(tp, &fold->res);
2276                 /* Caller holds reference to fold, so refcnt is always > 0
2277                  * after this.
2278                  */
2279                 refcount_dec(&fold->refcnt);
2280                 __fl_put(fold);
2281         } else {
2282                 if (handle) {
2283                         /* user specifies a handle and it doesn't exist */
2284                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2285                                             handle, GFP_ATOMIC);
2286
2287                         /* Filter with specified handle was concurrently
2288                          * inserted after initial check in cls_api. This is not
2289                          * necessarily an error if NLM_F_EXCL is not set in
2290                          * message flags. Returning EAGAIN will cause cls_api to
2291                          * try to update concurrently inserted rule.
2292                          */
2293                         if (err == -ENOSPC)
2294                                 err = -EAGAIN;
2295                 } else {
2296                         handle = 1;
2297                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2298                                             INT_MAX, GFP_ATOMIC);
2299                 }
2300                 if (err)
2301                         goto errout_hw;
2302
2303                 refcount_inc(&fnew->refcnt);
2304                 fnew->handle = handle;
2305                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2306                 spin_unlock(&tp->lock);
2307         }
2308
2309         *arg = fnew;
2310
2311         kfree(tb);
2312         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2313         return 0;
2314
2315 errout_ht:
2316         spin_lock(&tp->lock);
2317 errout_hw:
2318         fnew->deleted = true;
2319         spin_unlock(&tp->lock);
2320         if (!tc_skip_hw(fnew->flags))
2321                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2322         if (in_ht)
2323                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2324                                        fnew->mask->filter_ht_params);
2325 errout_mask:
2326         fl_mask_put(head, fnew->mask);
2327 errout:
2328         __fl_put(fnew);
2329 errout_tb:
2330         kfree(tb);
2331 errout_mask_alloc:
2332         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2333 errout_fold:
2334         if (fold)
2335                 __fl_put(fold);
2336         return err;
2337 }
2338
2339 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2340                      bool rtnl_held, struct netlink_ext_ack *extack)
2341 {
2342         struct cls_fl_head *head = fl_head_dereference(tp);
2343         struct cls_fl_filter *f = arg;
2344         bool last_on_mask;
2345         int err = 0;
2346
2347         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2348         *last = list_empty(&head->masks);
2349         __fl_put(f);
2350
2351         return err;
2352 }
2353
2354 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2355                     bool rtnl_held)
2356 {
2357         struct cls_fl_head *head = fl_head_dereference(tp);
2358         unsigned long id = arg->cookie, tmp;
2359         struct cls_fl_filter *f;
2360
2361         arg->count = arg->skip;
2362
2363         rcu_read_lock();
2364         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2365                 /* don't return filters that are being deleted */
2366                 if (!refcount_inc_not_zero(&f->refcnt))
2367                         continue;
2368                 rcu_read_unlock();
2369
2370                 if (arg->fn(tp, f, arg) < 0) {
2371                         __fl_put(f);
2372                         arg->stop = 1;
2373                         rcu_read_lock();
2374                         break;
2375                 }
2376                 __fl_put(f);
2377                 arg->count++;
2378                 rcu_read_lock();
2379         }
2380         rcu_read_unlock();
2381         arg->cookie = id;
2382 }
2383
2384 static struct cls_fl_filter *
2385 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2386 {
2387         struct cls_fl_head *head = fl_head_dereference(tp);
2388
2389         spin_lock(&tp->lock);
2390         if (list_empty(&head->hw_filters)) {
2391                 spin_unlock(&tp->lock);
2392                 return NULL;
2393         }
2394
2395         if (!f)
2396                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2397                                hw_list);
2398         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2399                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2400                         spin_unlock(&tp->lock);
2401                         return f;
2402                 }
2403         }
2404
2405         spin_unlock(&tp->lock);
2406         return NULL;
2407 }
2408
2409 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2410                         void *cb_priv, struct netlink_ext_ack *extack)
2411 {
2412         struct tcf_block *block = tp->chain->block;
2413         struct flow_cls_offload cls_flower = {};
2414         struct cls_fl_filter *f = NULL;
2415         int err;
2416
2417         /* hw_filters list can only be changed by hw offload functions after
2418          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2419          * iterating it.
2420          */
2421         ASSERT_RTNL();
2422
2423         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2424                 cls_flower.rule =
2425                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2426                 if (!cls_flower.rule) {
2427                         __fl_put(f);
2428                         return -ENOMEM;
2429                 }
2430
2431                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2432                                            extack);
2433                 cls_flower.command = add ?
2434                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2435                 cls_flower.cookie = (unsigned long)f;
2436                 cls_flower.rule->match.dissector = &f->mask->dissector;
2437                 cls_flower.rule->match.mask = &f->mask->key;
2438                 cls_flower.rule->match.key = &f->mkey;
2439
2440                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
2441                                               cls_flower.common.extack);
2442                 if (err) {
2443                         kfree(cls_flower.rule);
2444                         if (tc_skip_sw(f->flags)) {
2445                                 __fl_put(f);
2446                                 return err;
2447                         }
2448                         goto next_flow;
2449                 }
2450
2451                 cls_flower.classid = f->res.classid;
2452
2453                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2454                                             TC_SETUP_CLSFLOWER, &cls_flower,
2455                                             cb_priv, &f->flags,
2456                                             &f->in_hw_count);
2457                 tc_cleanup_offload_action(&cls_flower.rule->action);
2458                 kfree(cls_flower.rule);
2459
2460                 if (err) {
2461                         __fl_put(f);
2462                         return err;
2463                 }
2464 next_flow:
2465                 __fl_put(f);
2466         }
2467
2468         return 0;
2469 }
2470
2471 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2472 {
2473         struct flow_cls_offload *cls_flower = type_data;
2474         struct cls_fl_filter *f =
2475                 (struct cls_fl_filter *) cls_flower->cookie;
2476         struct cls_fl_head *head = fl_head_dereference(tp);
2477
2478         spin_lock(&tp->lock);
2479         list_add(&f->hw_list, &head->hw_filters);
2480         spin_unlock(&tp->lock);
2481 }
2482
2483 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2484 {
2485         struct flow_cls_offload *cls_flower = type_data;
2486         struct cls_fl_filter *f =
2487                 (struct cls_fl_filter *) cls_flower->cookie;
2488
2489         spin_lock(&tp->lock);
2490         if (!list_empty(&f->hw_list))
2491                 list_del_init(&f->hw_list);
2492         spin_unlock(&tp->lock);
2493 }
2494
2495 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2496                               struct fl_flow_tmplt *tmplt)
2497 {
2498         struct flow_cls_offload cls_flower = {};
2499         struct tcf_block *block = chain->block;
2500
2501         cls_flower.rule = flow_rule_alloc(0);
2502         if (!cls_flower.rule)
2503                 return -ENOMEM;
2504
2505         cls_flower.common.chain_index = chain->index;
2506         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2507         cls_flower.cookie = (unsigned long) tmplt;
2508         cls_flower.rule->match.dissector = &tmplt->dissector;
2509         cls_flower.rule->match.mask = &tmplt->mask;
2510         cls_flower.rule->match.key = &tmplt->dummy_key;
2511
2512         /* We don't care if driver (any of them) fails to handle this
2513          * call. It serves just as a hint for it.
2514          */
2515         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2516         kfree(cls_flower.rule);
2517
2518         return 0;
2519 }
2520
2521 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2522                                 struct fl_flow_tmplt *tmplt)
2523 {
2524         struct flow_cls_offload cls_flower = {};
2525         struct tcf_block *block = chain->block;
2526
2527         cls_flower.common.chain_index = chain->index;
2528         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2529         cls_flower.cookie = (unsigned long) tmplt;
2530
2531         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2532 }
2533
2534 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2535                              struct nlattr **tca,
2536                              struct netlink_ext_ack *extack)
2537 {
2538         struct fl_flow_tmplt *tmplt;
2539         struct nlattr **tb;
2540         int err;
2541
2542         if (!tca[TCA_OPTIONS])
2543                 return ERR_PTR(-EINVAL);
2544
2545         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2546         if (!tb)
2547                 return ERR_PTR(-ENOBUFS);
2548         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2549                                           tca[TCA_OPTIONS], fl_policy, NULL);
2550         if (err)
2551                 goto errout_tb;
2552
2553         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2554         if (!tmplt) {
2555                 err = -ENOMEM;
2556                 goto errout_tb;
2557         }
2558         tmplt->chain = chain;
2559         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2560         if (err)
2561                 goto errout_tmplt;
2562
2563         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2564
2565         err = fl_hw_create_tmplt(chain, tmplt);
2566         if (err)
2567                 goto errout_tmplt;
2568
2569         kfree(tb);
2570         return tmplt;
2571
2572 errout_tmplt:
2573         kfree(tmplt);
2574 errout_tb:
2575         kfree(tb);
2576         return ERR_PTR(err);
2577 }
2578
2579 static void fl_tmplt_destroy(void *tmplt_priv)
2580 {
2581         struct fl_flow_tmplt *tmplt = tmplt_priv;
2582
2583         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2584         kfree(tmplt);
2585 }
2586
2587 static int fl_dump_key_val(struct sk_buff *skb,
2588                            void *val, int val_type,
2589                            void *mask, int mask_type, int len)
2590 {
2591         int err;
2592
2593         if (!memchr_inv(mask, 0, len))
2594                 return 0;
2595         err = nla_put(skb, val_type, len, val);
2596         if (err)
2597                 return err;
2598         if (mask_type != TCA_FLOWER_UNSPEC) {
2599                 err = nla_put(skb, mask_type, len, mask);
2600                 if (err)
2601                         return err;
2602         }
2603         return 0;
2604 }
2605
2606 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2607                                   struct fl_flow_key *mask)
2608 {
2609         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2610                             TCA_FLOWER_KEY_PORT_DST_MIN,
2611                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2612                             sizeof(key->tp_range.tp_min.dst)) ||
2613             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2614                             TCA_FLOWER_KEY_PORT_DST_MAX,
2615                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2616                             sizeof(key->tp_range.tp_max.dst)) ||
2617             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2618                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2619                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2620                             sizeof(key->tp_range.tp_min.src)) ||
2621             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2622                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2623                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2624                             sizeof(key->tp_range.tp_max.src)))
2625                 return -1;
2626
2627         return 0;
2628 }
2629
2630 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2631                                     struct flow_dissector_key_mpls *mpls_key,
2632                                     struct flow_dissector_key_mpls *mpls_mask,
2633                                     u8 lse_index)
2634 {
2635         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2636         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2637         int err;
2638
2639         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2640                          lse_index + 1);
2641         if (err)
2642                 return err;
2643
2644         if (lse_mask->mpls_ttl) {
2645                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2646                                  lse_key->mpls_ttl);
2647                 if (err)
2648                         return err;
2649         }
2650         if (lse_mask->mpls_bos) {
2651                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2652                                  lse_key->mpls_bos);
2653                 if (err)
2654                         return err;
2655         }
2656         if (lse_mask->mpls_tc) {
2657                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2658                                  lse_key->mpls_tc);
2659                 if (err)
2660                         return err;
2661         }
2662         if (lse_mask->mpls_label) {
2663                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2664                                   lse_key->mpls_label);
2665                 if (err)
2666                         return err;
2667         }
2668
2669         return 0;
2670 }
2671
2672 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2673                                  struct flow_dissector_key_mpls *mpls_key,
2674                                  struct flow_dissector_key_mpls *mpls_mask)
2675 {
2676         struct nlattr *opts;
2677         struct nlattr *lse;
2678         u8 lse_index;
2679         int err;
2680
2681         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2682         if (!opts)
2683                 return -EMSGSIZE;
2684
2685         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2686                 if (!(mpls_mask->used_lses & 1 << lse_index))
2687                         continue;
2688
2689                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2690                 if (!lse) {
2691                         err = -EMSGSIZE;
2692                         goto err_opts;
2693                 }
2694
2695                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2696                                                lse_index);
2697                 if (err)
2698                         goto err_opts_lse;
2699                 nla_nest_end(skb, lse);
2700         }
2701         nla_nest_end(skb, opts);
2702
2703         return 0;
2704
2705 err_opts_lse:
2706         nla_nest_cancel(skb, lse);
2707 err_opts:
2708         nla_nest_cancel(skb, opts);
2709
2710         return err;
2711 }
2712
2713 static int fl_dump_key_mpls(struct sk_buff *skb,
2714                             struct flow_dissector_key_mpls *mpls_key,
2715                             struct flow_dissector_key_mpls *mpls_mask)
2716 {
2717         struct flow_dissector_mpls_lse *lse_mask;
2718         struct flow_dissector_mpls_lse *lse_key;
2719         int err;
2720
2721         if (!mpls_mask->used_lses)
2722                 return 0;
2723
2724         lse_mask = &mpls_mask->ls[0];
2725         lse_key = &mpls_key->ls[0];
2726
2727         /* For backward compatibility, don't use the MPLS nested attributes if
2728          * the rule can be expressed using the old attributes.
2729          */
2730         if (mpls_mask->used_lses & ~1 ||
2731             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2732              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2733                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2734
2735         if (lse_mask->mpls_ttl) {
2736                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2737                                  lse_key->mpls_ttl);
2738                 if (err)
2739                         return err;
2740         }
2741         if (lse_mask->mpls_tc) {
2742                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2743                                  lse_key->mpls_tc);
2744                 if (err)
2745                         return err;
2746         }
2747         if (lse_mask->mpls_label) {
2748                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2749                                   lse_key->mpls_label);
2750                 if (err)
2751                         return err;
2752         }
2753         if (lse_mask->mpls_bos) {
2754                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2755                                  lse_key->mpls_bos);
2756                 if (err)
2757                         return err;
2758         }
2759         return 0;
2760 }
2761
2762 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2763                           struct flow_dissector_key_ip *key,
2764                           struct flow_dissector_key_ip *mask)
2765 {
2766         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2767         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2768         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2769         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2770
2771         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2772             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2773                 return -1;
2774
2775         return 0;
2776 }
2777
2778 static int fl_dump_key_vlan(struct sk_buff *skb,
2779                             int vlan_id_key, int vlan_prio_key,
2780                             struct flow_dissector_key_vlan *vlan_key,
2781                             struct flow_dissector_key_vlan *vlan_mask)
2782 {
2783         int err;
2784
2785         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2786                 return 0;
2787         if (vlan_mask->vlan_id) {
2788                 err = nla_put_u16(skb, vlan_id_key,
2789                                   vlan_key->vlan_id);
2790                 if (err)
2791                         return err;
2792         }
2793         if (vlan_mask->vlan_priority) {
2794                 err = nla_put_u8(skb, vlan_prio_key,
2795                                  vlan_key->vlan_priority);
2796                 if (err)
2797                         return err;
2798         }
2799         return 0;
2800 }
2801
2802 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2803                             u32 *flower_key, u32 *flower_mask,
2804                             u32 flower_flag_bit, u32 dissector_flag_bit)
2805 {
2806         if (dissector_mask & dissector_flag_bit) {
2807                 *flower_mask |= flower_flag_bit;
2808                 if (dissector_key & dissector_flag_bit)
2809                         *flower_key |= flower_flag_bit;
2810         }
2811 }
2812
2813 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2814 {
2815         u32 key, mask;
2816         __be32 _key, _mask;
2817         int err;
2818
2819         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2820                 return 0;
2821
2822         key = 0;
2823         mask = 0;
2824
2825         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2826                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2827         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2828                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2829                         FLOW_DIS_FIRST_FRAG);
2830
2831         _key = cpu_to_be32(key);
2832         _mask = cpu_to_be32(mask);
2833
2834         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2835         if (err)
2836                 return err;
2837
2838         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2839 }
2840
2841 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2842                                   struct flow_dissector_key_enc_opts *enc_opts)
2843 {
2844         struct geneve_opt *opt;
2845         struct nlattr *nest;
2846         int opt_off = 0;
2847
2848         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2849         if (!nest)
2850                 goto nla_put_failure;
2851
2852         while (enc_opts->len > opt_off) {
2853                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2854
2855                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2856                                  opt->opt_class))
2857                         goto nla_put_failure;
2858                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2859                                opt->type))
2860                         goto nla_put_failure;
2861                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2862                             opt->length * 4, opt->opt_data))
2863                         goto nla_put_failure;
2864
2865                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2866         }
2867         nla_nest_end(skb, nest);
2868         return 0;
2869
2870 nla_put_failure:
2871         nla_nest_cancel(skb, nest);
2872         return -EMSGSIZE;
2873 }
2874
2875 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2876                                  struct flow_dissector_key_enc_opts *enc_opts)
2877 {
2878         struct vxlan_metadata *md;
2879         struct nlattr *nest;
2880
2881         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2882         if (!nest)
2883                 goto nla_put_failure;
2884
2885         md = (struct vxlan_metadata *)&enc_opts->data[0];
2886         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2887                 goto nla_put_failure;
2888
2889         nla_nest_end(skb, nest);
2890         return 0;
2891
2892 nla_put_failure:
2893         nla_nest_cancel(skb, nest);
2894         return -EMSGSIZE;
2895 }
2896
2897 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2898                                   struct flow_dissector_key_enc_opts *enc_opts)
2899 {
2900         struct erspan_metadata *md;
2901         struct nlattr *nest;
2902
2903         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2904         if (!nest)
2905                 goto nla_put_failure;
2906
2907         md = (struct erspan_metadata *)&enc_opts->data[0];
2908         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2909                 goto nla_put_failure;
2910
2911         if (md->version == 1 &&
2912             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2913                 goto nla_put_failure;
2914
2915         if (md->version == 2 &&
2916             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2917                         md->u.md2.dir) ||
2918              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2919                         get_hwid(&md->u.md2))))
2920                 goto nla_put_failure;
2921
2922         nla_nest_end(skb, nest);
2923         return 0;
2924
2925 nla_put_failure:
2926         nla_nest_cancel(skb, nest);
2927         return -EMSGSIZE;
2928 }
2929
2930 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
2931                                struct flow_dissector_key_enc_opts *enc_opts)
2932
2933 {
2934         struct gtp_pdu_session_info *session_info;
2935         struct nlattr *nest;
2936
2937         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
2938         if (!nest)
2939                 goto nla_put_failure;
2940
2941         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
2942
2943         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
2944                        session_info->pdu_type))
2945                 goto nla_put_failure;
2946
2947         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
2948                 goto nla_put_failure;
2949
2950         nla_nest_end(skb, nest);
2951         return 0;
2952
2953 nla_put_failure:
2954         nla_nest_cancel(skb, nest);
2955         return -EMSGSIZE;
2956 }
2957
2958 static int fl_dump_key_ct(struct sk_buff *skb,
2959                           struct flow_dissector_key_ct *key,
2960                           struct flow_dissector_key_ct *mask)
2961 {
2962         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2963             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2964                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2965                             sizeof(key->ct_state)))
2966                 goto nla_put_failure;
2967
2968         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2969             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2970                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2971                             sizeof(key->ct_zone)))
2972                 goto nla_put_failure;
2973
2974         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2975             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2976                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2977                             sizeof(key->ct_mark)))
2978                 goto nla_put_failure;
2979
2980         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2981             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2982                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2983                             sizeof(key->ct_labels)))
2984                 goto nla_put_failure;
2985
2986         return 0;
2987
2988 nla_put_failure:
2989         return -EMSGSIZE;
2990 }
2991
2992 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2993                                struct flow_dissector_key_enc_opts *enc_opts)
2994 {
2995         struct nlattr *nest;
2996         int err;
2997
2998         if (!enc_opts->len)
2999                 return 0;
3000
3001         nest = nla_nest_start_noflag(skb, enc_opt_type);
3002         if (!nest)
3003                 goto nla_put_failure;
3004
3005         switch (enc_opts->dst_opt_type) {
3006         case TUNNEL_GENEVE_OPT:
3007                 err = fl_dump_key_geneve_opt(skb, enc_opts);
3008                 if (err)
3009                         goto nla_put_failure;
3010                 break;
3011         case TUNNEL_VXLAN_OPT:
3012                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
3013                 if (err)
3014                         goto nla_put_failure;
3015                 break;
3016         case TUNNEL_ERSPAN_OPT:
3017                 err = fl_dump_key_erspan_opt(skb, enc_opts);
3018                 if (err)
3019                         goto nla_put_failure;
3020                 break;
3021         case TUNNEL_GTP_OPT:
3022                 err = fl_dump_key_gtp_opt(skb, enc_opts);
3023                 if (err)
3024                         goto nla_put_failure;
3025                 break;
3026         default:
3027                 goto nla_put_failure;
3028         }
3029         nla_nest_end(skb, nest);
3030         return 0;
3031
3032 nla_put_failure:
3033         nla_nest_cancel(skb, nest);
3034         return -EMSGSIZE;
3035 }
3036
3037 static int fl_dump_key_enc_opt(struct sk_buff *skb,
3038                                struct flow_dissector_key_enc_opts *key_opts,
3039                                struct flow_dissector_key_enc_opts *msk_opts)
3040 {
3041         int err;
3042
3043         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
3044         if (err)
3045                 return err;
3046
3047         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
3048 }
3049
3050 static int fl_dump_key(struct sk_buff *skb, struct net *net,
3051                        struct fl_flow_key *key, struct fl_flow_key *mask)
3052 {
3053         if (mask->meta.ingress_ifindex) {
3054                 struct net_device *dev;
3055
3056                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
3057                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
3058                         goto nla_put_failure;
3059         }
3060
3061         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
3062                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
3063                             sizeof(key->eth.dst)) ||
3064             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
3065                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
3066                             sizeof(key->eth.src)) ||
3067             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
3068                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
3069                             sizeof(key->basic.n_proto)))
3070                 goto nla_put_failure;
3071
3072         if (mask->num_of_vlans.num_of_vlans) {
3073                 if (nla_put_u8(skb, TCA_FLOWER_KEY_NUM_OF_VLANS, key->num_of_vlans.num_of_vlans))
3074                         goto nla_put_failure;
3075         }
3076
3077         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
3078                 goto nla_put_failure;
3079
3080         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
3081                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
3082                 goto nla_put_failure;
3083
3084         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
3085                              TCA_FLOWER_KEY_CVLAN_PRIO,
3086                              &key->cvlan, &mask->cvlan) ||
3087             (mask->cvlan.vlan_tpid &&
3088              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3089                           key->cvlan.vlan_tpid)))
3090                 goto nla_put_failure;
3091
3092         if (mask->basic.n_proto) {
3093                 if (mask->cvlan.vlan_eth_type) {
3094                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3095                                          key->basic.n_proto))
3096                                 goto nla_put_failure;
3097                 } else if (mask->vlan.vlan_eth_type) {
3098                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3099                                          key->vlan.vlan_eth_type))
3100                                 goto nla_put_failure;
3101                 }
3102         }
3103
3104         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3105              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3106             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3107                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3108                             sizeof(key->basic.ip_proto)) ||
3109             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3110                 goto nla_put_failure;
3111
3112         if (mask->pppoe.session_id) {
3113                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPPOE_SID,
3114                                  key->pppoe.session_id))
3115                         goto nla_put_failure;
3116         }
3117         if (mask->basic.n_proto && mask->pppoe.ppp_proto) {
3118                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPP_PROTO,
3119                                  key->pppoe.ppp_proto))
3120                         goto nla_put_failure;
3121         }
3122
3123         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3124             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3125                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3126                              sizeof(key->ipv4.src)) ||
3127              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3128                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3129                              sizeof(key->ipv4.dst))))
3130                 goto nla_put_failure;
3131         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3132                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3133                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3134                                   sizeof(key->ipv6.src)) ||
3135                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3136                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3137                                   sizeof(key->ipv6.dst))))
3138                 goto nla_put_failure;
3139
3140         if (key->basic.ip_proto == IPPROTO_TCP &&
3141             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3142                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3143                              sizeof(key->tp.src)) ||
3144              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3145                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3146                              sizeof(key->tp.dst)) ||
3147              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3148                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3149                              sizeof(key->tcp.flags))))
3150                 goto nla_put_failure;
3151         else if (key->basic.ip_proto == IPPROTO_UDP &&
3152                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3153                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3154                                   sizeof(key->tp.src)) ||
3155                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3156                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3157                                   sizeof(key->tp.dst))))
3158                 goto nla_put_failure;
3159         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3160                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3161                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3162                                   sizeof(key->tp.src)) ||
3163                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3164                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3165                                   sizeof(key->tp.dst))))
3166                 goto nla_put_failure;
3167         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3168                  key->basic.ip_proto == IPPROTO_ICMP &&
3169                  (fl_dump_key_val(skb, &key->icmp.type,
3170                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3171                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3172                                   sizeof(key->icmp.type)) ||
3173                   fl_dump_key_val(skb, &key->icmp.code,
3174                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3175                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3176                                   sizeof(key->icmp.code))))
3177                 goto nla_put_failure;
3178         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3179                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3180                  (fl_dump_key_val(skb, &key->icmp.type,
3181                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3182                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3183                                   sizeof(key->icmp.type)) ||
3184                   fl_dump_key_val(skb, &key->icmp.code,
3185                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3186                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3187                                   sizeof(key->icmp.code))))
3188                 goto nla_put_failure;
3189         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3190                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3191                  (fl_dump_key_val(skb, &key->arp.sip,
3192                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3193                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3194                                   sizeof(key->arp.sip)) ||
3195                   fl_dump_key_val(skb, &key->arp.tip,
3196                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3197                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3198                                   sizeof(key->arp.tip)) ||
3199                   fl_dump_key_val(skb, &key->arp.op,
3200                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3201                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3202                                   sizeof(key->arp.op)) ||
3203                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3204                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3205                                   sizeof(key->arp.sha)) ||
3206                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3207                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3208                                   sizeof(key->arp.tha))))
3209                 goto nla_put_failure;
3210         else if (key->basic.ip_proto == IPPROTO_L2TP &&
3211                  fl_dump_key_val(skb, &key->l2tpv3.session_id,
3212                                  TCA_FLOWER_KEY_L2TPV3_SID,
3213                                  &mask->l2tpv3.session_id,
3214                                  TCA_FLOWER_UNSPEC,
3215                                  sizeof(key->l2tpv3.session_id)))
3216                 goto nla_put_failure;
3217
3218         if ((key->basic.ip_proto == IPPROTO_TCP ||
3219              key->basic.ip_proto == IPPROTO_UDP ||
3220              key->basic.ip_proto == IPPROTO_SCTP) &&
3221              fl_dump_key_port_range(skb, key, mask))
3222                 goto nla_put_failure;
3223
3224         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3225             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3226                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3227                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3228                             sizeof(key->enc_ipv4.src)) ||
3229              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3230                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3231                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3232                              sizeof(key->enc_ipv4.dst))))
3233                 goto nla_put_failure;
3234         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3235                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3236                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3237                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3238                             sizeof(key->enc_ipv6.src)) ||
3239                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3240                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3241                                  &mask->enc_ipv6.dst,
3242                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3243                             sizeof(key->enc_ipv6.dst))))
3244                 goto nla_put_failure;
3245
3246         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3247                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3248                             sizeof(key->enc_key_id)) ||
3249             fl_dump_key_val(skb, &key->enc_tp.src,
3250                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3251                             &mask->enc_tp.src,
3252                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3253                             sizeof(key->enc_tp.src)) ||
3254             fl_dump_key_val(skb, &key->enc_tp.dst,
3255                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3256                             &mask->enc_tp.dst,
3257                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3258                             sizeof(key->enc_tp.dst)) ||
3259             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3260             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3261                 goto nla_put_failure;
3262
3263         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3264                 goto nla_put_failure;
3265
3266         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3267                 goto nla_put_failure;
3268
3269         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3270                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3271                              sizeof(key->hash.hash)))
3272                 goto nla_put_failure;
3273
3274         return 0;
3275
3276 nla_put_failure:
3277         return -EMSGSIZE;
3278 }
3279
3280 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3281                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3282 {
3283         struct cls_fl_filter *f = fh;
3284         struct nlattr *nest;
3285         struct fl_flow_key *key, *mask;
3286         bool skip_hw;
3287
3288         if (!f)
3289                 return skb->len;
3290
3291         t->tcm_handle = f->handle;
3292
3293         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3294         if (!nest)
3295                 goto nla_put_failure;
3296
3297         spin_lock(&tp->lock);
3298
3299         if (f->res.classid &&
3300             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3301                 goto nla_put_failure_locked;
3302
3303         key = &f->key;
3304         mask = &f->mask->key;
3305         skip_hw = tc_skip_hw(f->flags);
3306
3307         if (fl_dump_key(skb, net, key, mask))
3308                 goto nla_put_failure_locked;
3309
3310         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3311                 goto nla_put_failure_locked;
3312
3313         spin_unlock(&tp->lock);
3314
3315         if (!skip_hw)
3316                 fl_hw_update_stats(tp, f, rtnl_held);
3317
3318         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3319                 goto nla_put_failure;
3320
3321         if (tcf_exts_dump(skb, &f->exts))
3322                 goto nla_put_failure;
3323
3324         nla_nest_end(skb, nest);
3325
3326         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3327                 goto nla_put_failure;
3328
3329         return skb->len;
3330
3331 nla_put_failure_locked:
3332         spin_unlock(&tp->lock);
3333 nla_put_failure:
3334         nla_nest_cancel(skb, nest);
3335         return -1;
3336 }
3337
3338 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3339                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3340 {
3341         struct cls_fl_filter *f = fh;
3342         struct nlattr *nest;
3343         bool skip_hw;
3344
3345         if (!f)
3346                 return skb->len;
3347
3348         t->tcm_handle = f->handle;
3349
3350         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3351         if (!nest)
3352                 goto nla_put_failure;
3353
3354         spin_lock(&tp->lock);
3355
3356         skip_hw = tc_skip_hw(f->flags);
3357
3358         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3359                 goto nla_put_failure_locked;
3360
3361         spin_unlock(&tp->lock);
3362
3363         if (!skip_hw)
3364                 fl_hw_update_stats(tp, f, rtnl_held);
3365
3366         if (tcf_exts_terse_dump(skb, &f->exts))
3367                 goto nla_put_failure;
3368
3369         nla_nest_end(skb, nest);
3370
3371         return skb->len;
3372
3373 nla_put_failure_locked:
3374         spin_unlock(&tp->lock);
3375 nla_put_failure:
3376         nla_nest_cancel(skb, nest);
3377         return -1;
3378 }
3379
3380 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3381 {
3382         struct fl_flow_tmplt *tmplt = tmplt_priv;
3383         struct fl_flow_key *key, *mask;
3384         struct nlattr *nest;
3385
3386         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3387         if (!nest)
3388                 goto nla_put_failure;
3389
3390         key = &tmplt->dummy_key;
3391         mask = &tmplt->mask;
3392
3393         if (fl_dump_key(skb, net, key, mask))
3394                 goto nla_put_failure;
3395
3396         nla_nest_end(skb, nest);
3397
3398         return skb->len;
3399
3400 nla_put_failure:
3401         nla_nest_cancel(skb, nest);
3402         return -EMSGSIZE;
3403 }
3404
3405 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3406                           unsigned long base)
3407 {
3408         struct cls_fl_filter *f = fh;
3409
3410         tc_cls_bind_class(classid, cl, q, &f->res, base);
3411 }
3412
3413 static bool fl_delete_empty(struct tcf_proto *tp)
3414 {
3415         struct cls_fl_head *head = fl_head_dereference(tp);
3416
3417         spin_lock(&tp->lock);
3418         tp->deleting = idr_is_empty(&head->handle_idr);
3419         spin_unlock(&tp->lock);
3420
3421         return tp->deleting;
3422 }
3423
3424 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3425         .kind           = "flower",
3426         .classify       = fl_classify,
3427         .init           = fl_init,
3428         .destroy        = fl_destroy,
3429         .get            = fl_get,
3430         .put            = fl_put,
3431         .change         = fl_change,
3432         .delete         = fl_delete,
3433         .delete_empty   = fl_delete_empty,
3434         .walk           = fl_walk,
3435         .reoffload      = fl_reoffload,
3436         .hw_add         = fl_hw_add,
3437         .hw_del         = fl_hw_del,
3438         .dump           = fl_dump,
3439         .terse_dump     = fl_terse_dump,
3440         .bind_class     = fl_bind_class,
3441         .tmplt_create   = fl_tmplt_create,
3442         .tmplt_destroy  = fl_tmplt_destroy,
3443         .tmplt_dump     = fl_tmplt_dump,
3444         .owner          = THIS_MODULE,
3445         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3446 };
3447
3448 static int __init cls_fl_init(void)
3449 {
3450         return register_tcf_proto_ops(&cls_fl_ops);
3451 }
3452
3453 static void __exit cls_fl_exit(void)
3454 {
3455         unregister_tcf_proto_ops(&cls_fl_ops);
3456 }
3457
3458 module_init(cls_fl_init);
3459 module_exit(cls_fl_exit);
3460
3461 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3462 MODULE_DESCRIPTION("Flower classifier");
3463 MODULE_LICENSE("GPL v2");