net/sched/cls_flower.c

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