net/sched/cls_flower.c

   1 /*
   2  * net/sched/cls_flower.c               Flower classifier
   3  *
   4  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/init.h>
  14 #include <linux/module.h>
  15 #include <linux/rhashtable.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/refcount.h>
  18
  19 #include <linux/if_ether.h>
  20 #include <linux/in6.h>
  21 #include <linux/ip.h>
  22 #include <linux/mpls.h>
  23
  24 #include <net/sch_generic.h>
  25 #include <net/pkt_cls.h>
  26 #include <net/ip.h>
  27 #include <net/flow_dissector.h>
  28 #include <net/geneve.h>
  29
  30 #include <net/dst.h>
  31 #include <net/dst_metadata.h>
  32
  33 struct fl_flow_key {
  34         int     indev_ifindex;
  35         struct flow_dissector_key_control control;
  36         struct flow_dissector_key_control enc_control;
  37         struct flow_dissector_key_basic basic;
  38         struct flow_dissector_key_eth_addrs eth;
  39         struct flow_dissector_key_vlan vlan;
  40         struct flow_dissector_key_vlan cvlan;
  41         union {
  42                 struct flow_dissector_key_ipv4_addrs ipv4;
  43                 struct flow_dissector_key_ipv6_addrs ipv6;
  44         };
  45         struct flow_dissector_key_ports tp;
  46         struct flow_dissector_key_icmp icmp;
  47         struct flow_dissector_key_arp arp;
  48         struct flow_dissector_key_keyid enc_key_id;
  49         union {
  50                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  51                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  52         };
  53         struct flow_dissector_key_ports enc_tp;
  54         struct flow_dissector_key_mpls mpls;
  55         struct flow_dissector_key_tcp tcp;
  56         struct flow_dissector_key_ip ip;
  57         struct flow_dissector_key_ip enc_ip;
  58         struct flow_dissector_key_enc_opts enc_opts;
  59         struct flow_dissector_key_ports tp_min;
  60         struct flow_dissector_key_ports tp_max;
  61 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  62
  63 struct fl_flow_mask_range {
  64         unsigned short int start;
  65         unsigned short int end;
  66 };
  67
  68 struct fl_flow_mask {
  69         struct fl_flow_key key;
  70         struct fl_flow_mask_range range;
  71         u32 flags;
  72         struct rhash_head ht_node;
  73         struct rhashtable ht;
  74         struct rhashtable_params filter_ht_params;
  75         struct flow_dissector dissector;
  76         struct list_head filters;
  77         struct rcu_work rwork;
  78         struct list_head list;
  79         refcount_t refcnt;
  80 };
  81
  82 struct fl_flow_tmplt {
  83         struct fl_flow_key dummy_key;
  84         struct fl_flow_key mask;
  85         struct flow_dissector dissector;
  86         struct tcf_chain *chain;
  87 };
  88
  89 struct cls_fl_head {
  90         struct rhashtable ht;
  91         spinlock_t masks_lock; /* Protect masks list */
  92         struct list_head masks;
  93         struct list_head hw_filters;
  94         struct rcu_work rwork;
  95         struct idr handle_idr;
  96 };
  97
  98 struct cls_fl_filter {
  99         struct fl_flow_mask *mask;
 100         struct rhash_head ht_node;
 101         struct fl_flow_key mkey;
 102         struct tcf_exts exts;
 103         struct tcf_result res;
 104         struct fl_flow_key key;
 105         struct list_head list;
 106         struct list_head hw_list;
 107         u32 handle;
 108         u32 flags;
 109         u32 in_hw_count;
 110         struct rcu_work rwork;
 111         struct net_device *hw_dev;
 112         /* Flower classifier is unlocked, which means that its reference counter
 113          * can be changed concurrently without any kind of external
 114          * synchronization. Use atomic reference counter to be concurrency-safe.
 115          */
 116         refcount_t refcnt;
 117         bool deleted;
 118 };
 119
 120 static const struct rhashtable_params mask_ht_params = {
 121         .key_offset = offsetof(struct fl_flow_mask, key),
 122         .key_len = sizeof(struct fl_flow_key),
 123         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 124         .automatic_shrinking = true,
 125 };
 126
 127 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 128 {
 129         return mask->range.end - mask->range.start;
 130 }
 131
 132 static void fl_mask_update_range(struct fl_flow_mask *mask)
 133 {
 134         const u8 *bytes = (const u8 *) &mask->key;
 135         size_t size = sizeof(mask->key);
 136         size_t i, first = 0, last;
 137
 138         for (i = 0; i < size; i++) {
 139                 if (bytes[i]) {
 140                         first = i;
 141                         break;
 142                 }
 143         }
 144         last = first;
 145         for (i = size - 1; i != first; i--) {
 146                 if (bytes[i]) {
 147                         last = i;
 148                         break;
 149                 }
 150         }
 151         mask->range.start = rounddown(first, sizeof(long));
 152         mask->range.end = roundup(last + 1, sizeof(long));
 153 }
 154
 155 static void *fl_key_get_start(struct fl_flow_key *key,
 156                               const struct fl_flow_mask *mask)
 157 {
 158         return (u8 *) key + mask->range.start;
 159 }
 160
 161 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 162                               struct fl_flow_mask *mask)
 163 {
 164         const long *lkey = fl_key_get_start(key, mask);
 165         const long *lmask = fl_key_get_start(&mask->key, mask);
 166         long *lmkey = fl_key_get_start(mkey, mask);
 167         int i;
 168
 169         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 170                 *lmkey++ = *lkey++ & *lmask++;
 171 }
 172
 173 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 174                                struct fl_flow_mask *mask)
 175 {
 176         const long *lmask = fl_key_get_start(&mask->key, mask);
 177         const long *ltmplt;
 178         int i;
 179
 180         if (!tmplt)
 181                 return true;
 182         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 183         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 184                 if (~*ltmplt++ & *lmask++)
 185                         return false;
 186         }
 187         return true;
 188 }
 189
 190 static void fl_clear_masked_range(struct fl_flow_key *key,
 191                                   struct fl_flow_mask *mask)
 192 {
 193         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 194 }
 195
 196 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 197                                   struct fl_flow_key *key,
 198                                   struct fl_flow_key *mkey)
 199 {
 200         __be16 min_mask, max_mask, min_val, max_val;
 201
 202         min_mask = htons(filter->mask->key.tp_min.dst);
 203         max_mask = htons(filter->mask->key.tp_max.dst);
 204         min_val = htons(filter->key.tp_min.dst);
 205         max_val = htons(filter->key.tp_max.dst);
 206
 207         if (min_mask && max_mask) {
 208                 if (htons(key->tp.dst) < min_val ||
 209                     htons(key->tp.dst) > max_val)
 210                         return false;
 211
 212                 /* skb does not have min and max values */
 213                 mkey->tp_min.dst = filter->mkey.tp_min.dst;
 214                 mkey->tp_max.dst = filter->mkey.tp_max.dst;
 215         }
 216         return true;
 217 }
 218
 219 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 220                                   struct fl_flow_key *key,
 221                                   struct fl_flow_key *mkey)
 222 {
 223         __be16 min_mask, max_mask, min_val, max_val;
 224
 225         min_mask = htons(filter->mask->key.tp_min.src);
 226         max_mask = htons(filter->mask->key.tp_max.src);
 227         min_val = htons(filter->key.tp_min.src);
 228         max_val = htons(filter->key.tp_max.src);
 229
 230         if (min_mask && max_mask) {
 231                 if (htons(key->tp.src) < min_val ||
 232                     htons(key->tp.src) > max_val)
 233                         return false;
 234
 235                 /* skb does not have min and max values */
 236                 mkey->tp_min.src = filter->mkey.tp_min.src;
 237                 mkey->tp_max.src = filter->mkey.tp_max.src;
 238         }
 239         return true;
 240 }
 241
 242 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 243                                          struct fl_flow_key *mkey)
 244 {
 245         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 246                                       mask->filter_ht_params);
 247 }
 248
 249 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 250                                              struct fl_flow_key *mkey,
 251                                              struct fl_flow_key *key)
 252 {
 253         struct cls_fl_filter *filter, *f;
 254
 255         list_for_each_entry_rcu(filter, &mask->filters, list) {
 256                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 257                         continue;
 258
 259                 if (!fl_range_port_src_cmp(filter, key, mkey))
 260                         continue;
 261
 262                 f = __fl_lookup(mask, mkey);
 263                 if (f)
 264                         return f;
 265         }
 266         return NULL;
 267 }
 268
 269 static struct cls_fl_filter *fl_lookup(struct fl_flow_mask *mask,
 270                                        struct fl_flow_key *mkey,
 271                                        struct fl_flow_key *key)
 272 {
 273         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 274                 return fl_lookup_range(mask, mkey, key);
 275
 276         return __fl_lookup(mask, mkey);
 277 }
 278
 279 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 280                        struct tcf_result *res)
 281 {
 282         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 283         struct cls_fl_filter *f;
 284         struct fl_flow_mask *mask;
 285         struct fl_flow_key skb_key;
 286         struct fl_flow_key skb_mkey;
 287
 288         list_for_each_entry_rcu(mask, &head->masks, list) {
 289                 fl_clear_masked_range(&skb_key, mask);
 290
 291                 skb_key.indev_ifindex = skb->skb_iif;
 292                 /* skb_flow_dissect() does not set n_proto in case an unknown
 293                  * protocol, so do it rather here.
 294                  */
 295                 skb_key.basic.n_proto = skb->protocol;
 296                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 297                 skb_flow_dissect(skb, &mask->dissector, &skb_key, 0);
 298
 299                 fl_set_masked_key(&skb_mkey, &skb_key, mask);
 300
 301                 f = fl_lookup(mask, &skb_mkey, &skb_key);
 302                 if (f && !tc_skip_sw(f->flags)) {
 303                         *res = f->res;
 304                         return tcf_exts_exec(skb, &f->exts, res);
 305                 }
 306         }
 307         return -1;
 308 }
 309
 310 static int fl_init(struct tcf_proto *tp)
 311 {
 312         struct cls_fl_head *head;
 313
 314         head = kzalloc(sizeof(*head), GFP_KERNEL);
 315         if (!head)
 316                 return -ENOBUFS;
 317
 318         spin_lock_init(&head->masks_lock);
 319         INIT_LIST_HEAD_RCU(&head->masks);
 320         INIT_LIST_HEAD(&head->hw_filters);
 321         rcu_assign_pointer(tp->root, head);
 322         idr_init(&head->handle_idr);
 323
 324         return rhashtable_init(&head->ht, &mask_ht_params);
 325 }
 326
 327 static void fl_mask_free(struct fl_flow_mask *mask)
 328 {
 329         WARN_ON(!list_empty(&mask->filters));
 330         rhashtable_destroy(&mask->ht);
 331         kfree(mask);
 332 }
 333
 334 static void fl_mask_free_work(struct work_struct *work)
 335 {
 336         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 337                                                  struct fl_flow_mask, rwork);
 338
 339         fl_mask_free(mask);
 340 }
 341
 342 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 343 {
 344         if (!refcount_dec_and_test(&mask->refcnt))
 345                 return false;
 346
 347         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 348
 349         spin_lock(&head->masks_lock);
 350         list_del_rcu(&mask->list);
 351         spin_unlock(&head->masks_lock);
 352
 353         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 354
 355         return true;
 356 }
 357
 358 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 359 {
 360         /* Flower classifier only changes root pointer during init and destroy.
 361          * Users must obtain reference to tcf_proto instance before calling its
 362          * API, so tp->root pointer is protected from concurrent call to
 363          * fl_destroy() by reference counting.
 364          */
 365         return rcu_dereference_raw(tp->root);
 366 }
 367
 368 static void __fl_destroy_filter(struct cls_fl_filter *f)
 369 {
 370         tcf_exts_destroy(&f->exts);
 371         tcf_exts_put_net(&f->exts);
 372         kfree(f);
 373 }
 374
 375 static void fl_destroy_filter_work(struct work_struct *work)
 376 {
 377         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 378                                         struct cls_fl_filter, rwork);
 379
 380         __fl_destroy_filter(f);
 381 }
 382
 383 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 384                                  bool rtnl_held, struct netlink_ext_ack *extack)
 385 {
 386         struct tc_cls_flower_offload cls_flower = {};
 387         struct tcf_block *block = tp->chain->block;
 388
 389         if (!rtnl_held)
 390                 rtnl_lock();
 391
 392         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 393         cls_flower.command = TC_CLSFLOWER_DESTROY;
 394         cls_flower.cookie = (unsigned long) f;
 395
 396         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
 397         spin_lock(&tp->lock);
 398         list_del_init(&f->hw_list);
 399         tcf_block_offload_dec(block, &f->flags);
 400         spin_unlock(&tp->lock);
 401
 402         if (!rtnl_held)
 403                 rtnl_unlock();
 404 }
 405
 406 static int fl_hw_replace_filter(struct tcf_proto *tp,
 407                                 struct cls_fl_filter *f, bool rtnl_held,
 408                                 struct netlink_ext_ack *extack)
 409 {
 410         struct cls_fl_head *head = fl_head_dereference(tp);
 411         struct tc_cls_flower_offload cls_flower = {};
 412         struct tcf_block *block = tp->chain->block;
 413         bool skip_sw = tc_skip_sw(f->flags);
 414         int err = 0;
 415
 416         if (!rtnl_held)
 417                 rtnl_lock();
 418
 419         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 420         if (!cls_flower.rule) {
 421                 err = -ENOMEM;
 422                 goto errout;
 423         }
 424
 425         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 426         cls_flower.command = TC_CLSFLOWER_REPLACE;
 427         cls_flower.cookie = (unsigned long) f;
 428         cls_flower.rule->match.dissector = &f->mask->dissector;
 429         cls_flower.rule->match.mask = &f->mask->key;
 430         cls_flower.rule->match.key = &f->mkey;
 431         cls_flower.classid = f->res.classid;
 432
 433         err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
 434         if (err) {
 435                 kfree(cls_flower.rule);
 436                 if (skip_sw)
 437                         NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
 438                 else
 439                         err = 0;
 440                 goto errout;
 441         }
 442
 443         err = tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, skip_sw);
 444         kfree(cls_flower.rule);
 445
 446         if (err < 0) {
 447                 fl_hw_destroy_filter(tp, f, true, NULL);
 448                 goto errout;
 449         } else if (err > 0) {
 450                 f->in_hw_count = err;
 451                 err = 0;
 452                 spin_lock(&tp->lock);
 453                 tcf_block_offload_inc(block, &f->flags);
 454                 spin_unlock(&tp->lock);
 455         }
 456
 457         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW)) {
 458                 err = -EINVAL;
 459                 goto errout;
 460         }
 461
 462         spin_lock(&tp->lock);
 463         list_add(&f->hw_list, &head->hw_filters);
 464         spin_unlock(&tp->lock);
 465 errout:
 466         if (!rtnl_held)
 467                 rtnl_unlock();
 468
 469         return err;
 470 }
 471
 472 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 473                                bool rtnl_held)
 474 {
 475         struct tc_cls_flower_offload cls_flower = {};
 476         struct tcf_block *block = tp->chain->block;
 477
 478         if (!rtnl_held)
 479                 rtnl_lock();
 480
 481         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 482         cls_flower.command = TC_CLSFLOWER_STATS;
 483         cls_flower.cookie = (unsigned long) f;
 484         cls_flower.classid = f->res.classid;
 485
 486         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
 487
 488         tcf_exts_stats_update(&f->exts, cls_flower.stats.bytes,
 489                               cls_flower.stats.pkts,
 490                               cls_flower.stats.lastused);
 491
 492         if (!rtnl_held)
 493                 rtnl_unlock();
 494 }
 495
 496 static void __fl_put(struct cls_fl_filter *f)
 497 {
 498         if (!refcount_dec_and_test(&f->refcnt))
 499                 return;
 500
 501         if (tcf_exts_get_net(&f->exts))
 502                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 503         else
 504                 __fl_destroy_filter(f);
 505 }
 506
 507 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 508 {
 509         struct cls_fl_filter *f;
 510
 511         rcu_read_lock();
 512         f = idr_find(&head->handle_idr, handle);
 513         if (f && !refcount_inc_not_zero(&f->refcnt))
 514                 f = NULL;
 515         rcu_read_unlock();
 516
 517         return f;
 518 }
 519
 520 static struct cls_fl_filter *fl_get_next_filter(struct tcf_proto *tp,
 521                                                 unsigned long *handle)
 522 {
 523         struct cls_fl_head *head = fl_head_dereference(tp);
 524         struct cls_fl_filter *f;
 525
 526         rcu_read_lock();
 527         while ((f = idr_get_next_ul(&head->handle_idr, handle))) {
 528                 /* don't return filters that are being deleted */
 529                 if (refcount_inc_not_zero(&f->refcnt))
 530                         break;
 531                 ++(*handle);
 532         }
 533         rcu_read_unlock();
 534
 535         return f;
 536 }
 537
 538 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 539                        bool *last, bool rtnl_held,
 540                        struct netlink_ext_ack *extack)
 541 {
 542         struct cls_fl_head *head = fl_head_dereference(tp);
 543
 544         *last = false;
 545
 546         spin_lock(&tp->lock);
 547         if (f->deleted) {
 548                 spin_unlock(&tp->lock);
 549                 return -ENOENT;
 550         }
 551
 552         f->deleted = true;
 553         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 554                                f->mask->filter_ht_params);
 555         idr_remove(&head->handle_idr, f->handle);
 556         list_del_rcu(&f->list);
 557         spin_unlock(&tp->lock);
 558
 559         *last = fl_mask_put(head, f->mask);
 560         if (!tc_skip_hw(f->flags))
 561                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 562         tcf_unbind_filter(tp, &f->res);
 563         __fl_put(f);
 564
 565         return 0;
 566 }
 567
 568 static void fl_destroy_sleepable(struct work_struct *work)
 569 {
 570         struct cls_fl_head *head = container_of(to_rcu_work(work),
 571                                                 struct cls_fl_head,
 572                                                 rwork);
 573
 574         rhashtable_destroy(&head->ht);
 575         kfree(head);
 576         module_put(THIS_MODULE);
 577 }
 578
 579 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 580                        struct netlink_ext_ack *extack)
 581 {
 582         struct cls_fl_head *head = fl_head_dereference(tp);
 583         struct fl_flow_mask *mask, *next_mask;
 584         struct cls_fl_filter *f, *next;
 585         bool last;
 586
 587         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 588                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 589                         __fl_delete(tp, f, &last, rtnl_held, extack);
 590                         if (last)
 591                                 break;
 592                 }
 593         }
 594         idr_destroy(&head->handle_idr);
 595
 596         __module_get(THIS_MODULE);
 597         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 598 }
 599
 600 static void fl_put(struct tcf_proto *tp, void *arg)
 601 {
 602         struct cls_fl_filter *f = arg;
 603
 604         __fl_put(f);
 605 }
 606
 607 static void *fl_get(struct tcf_proto *tp, u32 handle)
 608 {
 609         struct cls_fl_head *head = fl_head_dereference(tp);
 610
 611         return __fl_get(head, handle);
 612 }
 613
 614 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 615         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 616         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 617         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 618                                             .len = IFNAMSIZ },
 619         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 620         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 621         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 622         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 623         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 624         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 625         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 626         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 627         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 628         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 629         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 630         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 631         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 632         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 633         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 634         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 635         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 636         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 637         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 638         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 639         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 640         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 641         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 642         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 643         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 644         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 645         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 646         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 647         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 648         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 649         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 650         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 660         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 661         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 662         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 663         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 664         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 665         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 666         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 670         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 671         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 672         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 673         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 674         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 675         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 676         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 677         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 678         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 679         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 680         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 681         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 682         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 683         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 684         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 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 };
 701
 702 static const struct nla_policy
 703 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 704         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 705 };
 706
 707 static const struct nla_policy
 708 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 709         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 710         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 711         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 712                                                        .len = 128 },
 713 };
 714
 715 static void fl_set_key_val(struct nlattr **tb,
 716                            void *val, int val_type,
 717                            void *mask, int mask_type, int len)
 718 {
 719         if (!tb[val_type])
 720                 return;
 721         memcpy(val, nla_data(tb[val_type]), len);
 722         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 723                 memset(mask, 0xff, len);
 724         else
 725                 memcpy(mask, nla_data(tb[mask_type]), len);
 726 }
 727
 728 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 729                                  struct fl_flow_key *mask)
 730 {
 731         fl_set_key_val(tb, &key->tp_min.dst,
 732                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_min.dst,
 733                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.dst));
 734         fl_set_key_val(tb, &key->tp_max.dst,
 735                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_max.dst,
 736                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.dst));
 737         fl_set_key_val(tb, &key->tp_min.src,
 738                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_min.src,
 739                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.src));
 740         fl_set_key_val(tb, &key->tp_max.src,
 741                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_max.src,
 742                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.src));
 743
 744         if ((mask->tp_min.dst && mask->tp_max.dst &&
 745              htons(key->tp_max.dst) <= htons(key->tp_min.dst)) ||
 746              (mask->tp_min.src && mask->tp_max.src &&
 747               htons(key->tp_max.src) <= htons(key->tp_min.src)))
 748                 return -EINVAL;
 749
 750         return 0;
 751 }
 752
 753 static int fl_set_key_mpls(struct nlattr **tb,
 754                            struct flow_dissector_key_mpls *key_val,
 755                            struct flow_dissector_key_mpls *key_mask)
 756 {
 757         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 758                 key_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 759                 key_mask->mpls_ttl = MPLS_TTL_MASK;
 760         }
 761         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 762                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 763
 764                 if (bos & ~MPLS_BOS_MASK)
 765                         return -EINVAL;
 766                 key_val->mpls_bos = bos;
 767                 key_mask->mpls_bos = MPLS_BOS_MASK;
 768         }
 769         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 770                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 771
 772                 if (tc & ~MPLS_TC_MASK)
 773                         return -EINVAL;
 774                 key_val->mpls_tc = tc;
 775                 key_mask->mpls_tc = MPLS_TC_MASK;
 776         }
 777         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 778                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 779
 780                 if (label & ~MPLS_LABEL_MASK)
 781                         return -EINVAL;
 782                 key_val->mpls_label = label;
 783                 key_mask->mpls_label = MPLS_LABEL_MASK;
 784         }
 785         return 0;
 786 }
 787
 788 static void fl_set_key_vlan(struct nlattr **tb,
 789                             __be16 ethertype,
 790                             int vlan_id_key, int vlan_prio_key,
 791                             struct flow_dissector_key_vlan *key_val,
 792                             struct flow_dissector_key_vlan *key_mask)
 793 {
 794 #define VLAN_PRIORITY_MASK      0x7
 795
 796         if (tb[vlan_id_key]) {
 797                 key_val->vlan_id =
 798                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
 799                 key_mask->vlan_id = VLAN_VID_MASK;
 800         }
 801         if (tb[vlan_prio_key]) {
 802                 key_val->vlan_priority =
 803                         nla_get_u8(tb[vlan_prio_key]) &
 804                         VLAN_PRIORITY_MASK;
 805                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 806         }
 807         key_val->vlan_tpid = ethertype;
 808         key_mask->vlan_tpid = cpu_to_be16(~0);
 809 }
 810
 811 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
 812                             u32 *dissector_key, u32 *dissector_mask,
 813                             u32 flower_flag_bit, u32 dissector_flag_bit)
 814 {
 815         if (flower_mask & flower_flag_bit) {
 816                 *dissector_mask |= dissector_flag_bit;
 817                 if (flower_key & flower_flag_bit)
 818                         *dissector_key |= dissector_flag_bit;
 819         }
 820 }
 821
 822 static int fl_set_key_flags(struct nlattr **tb,
 823                             u32 *flags_key, u32 *flags_mask)
 824 {
 825         u32 key, mask;
 826
 827         /* mask is mandatory for flags */
 828         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK])
 829                 return -EINVAL;
 830
 831         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
 832         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
 833
 834         *flags_key  = 0;
 835         *flags_mask = 0;
 836
 837         fl_set_key_flag(key, mask, flags_key, flags_mask,
 838                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
 839         fl_set_key_flag(key, mask, flags_key, flags_mask,
 840                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
 841                         FLOW_DIS_FIRST_FRAG);
 842
 843         return 0;
 844 }
 845
 846 static void fl_set_key_ip(struct nlattr **tb, bool encap,
 847                           struct flow_dissector_key_ip *key,
 848                           struct flow_dissector_key_ip *mask)
 849 {
 850         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
 851         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
 852         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
 853         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
 854
 855         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
 856         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
 857 }
 858
 859 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
 860                              int depth, int option_len,
 861                              struct netlink_ext_ack *extack)
 862 {
 863         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
 864         struct nlattr *class = NULL, *type = NULL, *data = NULL;
 865         struct geneve_opt *opt;
 866         int err, data_len = 0;
 867
 868         if (option_len > sizeof(struct geneve_opt))
 869                 data_len = option_len - sizeof(struct geneve_opt);
 870
 871         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
 872         memset(opt, 0xff, option_len);
 873         opt->length = data_len / 4;
 874         opt->r1 = 0;
 875         opt->r2 = 0;
 876         opt->r3 = 0;
 877
 878         /* If no mask has been prodived we assume an exact match. */
 879         if (!depth)
 880                 return sizeof(struct geneve_opt) + data_len;
 881
 882         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
 883                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
 884                 return -EINVAL;
 885         }
 886
 887         err = nla_parse_nested_deprecated(tb,
 888                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
 889                                           nla, geneve_opt_policy, extack);
 890         if (err < 0)
 891                 return err;
 892
 893         /* We are not allowed to omit any of CLASS, TYPE or DATA
 894          * fields from the key.
 895          */
 896         if (!option_len &&
 897             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
 898              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
 899              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
 900                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
 901                 return -EINVAL;
 902         }
 903
 904         /* Omitting any of CLASS, TYPE or DATA fields is allowed
 905          * for the mask.
 906          */
 907         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
 908                 int new_len = key->enc_opts.len;
 909
 910                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
 911                 data_len = nla_len(data);
 912                 if (data_len < 4) {
 913                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
 914                         return -ERANGE;
 915                 }
 916                 if (data_len % 4) {
 917                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
 918                         return -ERANGE;
 919                 }
 920
 921                 new_len += sizeof(struct geneve_opt) + data_len;
 922                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
 923                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
 924                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
 925                         return -ERANGE;
 926                 }
 927                 opt->length = data_len / 4;
 928                 memcpy(opt->opt_data, nla_data(data), data_len);
 929         }
 930
 931         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
 932                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
 933                 opt->opt_class = nla_get_be16(class);
 934         }
 935
 936         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
 937                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
 938                 opt->type = nla_get_u8(type);
 939         }
 940
 941         return sizeof(struct geneve_opt) + data_len;
 942 }
 943
 944 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
 945                           struct fl_flow_key *mask,
 946                           struct netlink_ext_ack *extack)
 947 {
 948         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
 949         int err, option_len, key_depth, msk_depth = 0;
 950
 951         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
 952                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
 953                                              enc_opts_policy, extack);
 954         if (err)
 955                 return err;
 956
 957         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
 958
 959         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
 960                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
 961                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
 962                                                      enc_opts_policy, extack);
 963                 if (err)
 964                         return err;
 965
 966                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 967                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 968         }
 969
 970         nla_for_each_attr(nla_opt_key, nla_enc_key,
 971                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
 972                 switch (nla_type(nla_opt_key)) {
 973                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
 974                         option_len = 0;
 975                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 976                         option_len = fl_set_geneve_opt(nla_opt_key, key,
 977                                                        key_depth, option_len,
 978                                                        extack);
 979                         if (option_len < 0)
 980                                 return option_len;
 981
 982                         key->enc_opts.len += option_len;
 983                         /* At the same time we need to parse through the mask
 984                          * in order to verify exact and mask attribute lengths.
 985                          */
 986                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 987                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
 988                                                        msk_depth, option_len,
 989                                                        extack);
 990                         if (option_len < 0)
 991                                 return option_len;
 992
 993                         mask->enc_opts.len += option_len;
 994                         if (key->enc_opts.len != mask->enc_opts.len) {
 995                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
 996                                 return -EINVAL;
 997                         }
 998
 999                         if (msk_depth)
1000                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1001                         break;
1002                 default:
1003                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1004                         return -EINVAL;
1005                 }
1006         }
1007
1008         return 0;
1009 }
1010
1011 static int fl_set_key(struct net *net, struct nlattr **tb,
1012                       struct fl_flow_key *key, struct fl_flow_key *mask,
1013                       struct netlink_ext_ack *extack)
1014 {
1015         __be16 ethertype;
1016         int ret = 0;
1017 #ifdef CONFIG_NET_CLS_IND
1018         if (tb[TCA_FLOWER_INDEV]) {
1019                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1020                 if (err < 0)
1021                         return err;
1022                 key->indev_ifindex = err;
1023                 mask->indev_ifindex = 0xffffffff;
1024         }
1025 #endif
1026
1027         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1028                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1029                        sizeof(key->eth.dst));
1030         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1031                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1032                        sizeof(key->eth.src));
1033
1034         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1035                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1036
1037                 if (eth_type_vlan(ethertype)) {
1038                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1039                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
1040                                         &mask->vlan);
1041
1042                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1043                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1044                                 if (eth_type_vlan(ethertype)) {
1045                                         fl_set_key_vlan(tb, ethertype,
1046                                                         TCA_FLOWER_KEY_CVLAN_ID,
1047                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1048                                                         &key->cvlan, &mask->cvlan);
1049                                         fl_set_key_val(tb, &key->basic.n_proto,
1050                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1051                                                        &mask->basic.n_proto,
1052                                                        TCA_FLOWER_UNSPEC,
1053                                                        sizeof(key->basic.n_proto));
1054                                 } else {
1055                                         key->basic.n_proto = ethertype;
1056                                         mask->basic.n_proto = cpu_to_be16(~0);
1057                                 }
1058                         }
1059                 } else {
1060                         key->basic.n_proto = ethertype;
1061                         mask->basic.n_proto = cpu_to_be16(~0);
1062                 }
1063         }
1064
1065         if (key->basic.n_proto == htons(ETH_P_IP) ||
1066             key->basic.n_proto == htons(ETH_P_IPV6)) {
1067                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1068                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1069                                sizeof(key->basic.ip_proto));
1070                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1071         }
1072
1073         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1074                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1075                 mask->control.addr_type = ~0;
1076                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1077                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1078                                sizeof(key->ipv4.src));
1079                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1080                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1081                                sizeof(key->ipv4.dst));
1082         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1083                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1084                 mask->control.addr_type = ~0;
1085                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1086                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1087                                sizeof(key->ipv6.src));
1088                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1089                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1090                                sizeof(key->ipv6.dst));
1091         }
1092
1093         if (key->basic.ip_proto == IPPROTO_TCP) {
1094                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1095                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1096                                sizeof(key->tp.src));
1097                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1098                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1099                                sizeof(key->tp.dst));
1100                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1101                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1102                                sizeof(key->tcp.flags));
1103         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1104                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1105                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1106                                sizeof(key->tp.src));
1107                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1108                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1109                                sizeof(key->tp.dst));
1110         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1111                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1112                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1113                                sizeof(key->tp.src));
1114                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1115                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1116                                sizeof(key->tp.dst));
1117         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1118                    key->basic.ip_proto == IPPROTO_ICMP) {
1119                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1120                                &mask->icmp.type,
1121                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1122                                sizeof(key->icmp.type));
1123                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1124                                &mask->icmp.code,
1125                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1126                                sizeof(key->icmp.code));
1127         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1128                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1129                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1130                                &mask->icmp.type,
1131                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1132                                sizeof(key->icmp.type));
1133                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1134                                &mask->icmp.code,
1135                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1136                                sizeof(key->icmp.code));
1137         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1138                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1139                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls);
1140                 if (ret)
1141                         return ret;
1142         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1143                    key->basic.n_proto == htons(ETH_P_RARP)) {
1144                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1145                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1146                                sizeof(key->arp.sip));
1147                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1148                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1149                                sizeof(key->arp.tip));
1150                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1151                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1152                                sizeof(key->arp.op));
1153                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1154                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1155                                sizeof(key->arp.sha));
1156                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1157                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1158                                sizeof(key->arp.tha));
1159         }
1160
1161         if (key->basic.ip_proto == IPPROTO_TCP ||
1162             key->basic.ip_proto == IPPROTO_UDP ||
1163             key->basic.ip_proto == IPPROTO_SCTP) {
1164                 ret = fl_set_key_port_range(tb, key, mask);
1165                 if (ret)
1166                         return ret;
1167         }
1168
1169         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1170             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1171                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1172                 mask->enc_control.addr_type = ~0;
1173                 fl_set_key_val(tb, &key->enc_ipv4.src,
1174                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1175                                &mask->enc_ipv4.src,
1176                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1177                                sizeof(key->enc_ipv4.src));
1178                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1179                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1180                                &mask->enc_ipv4.dst,
1181                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1182                                sizeof(key->enc_ipv4.dst));
1183         }
1184
1185         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1186             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1187                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1188                 mask->enc_control.addr_type = ~0;
1189                 fl_set_key_val(tb, &key->enc_ipv6.src,
1190                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1191                                &mask->enc_ipv6.src,
1192                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1193                                sizeof(key->enc_ipv6.src));
1194                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1195                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1196                                &mask->enc_ipv6.dst,
1197                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1198                                sizeof(key->enc_ipv6.dst));
1199         }
1200
1201         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1202                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1203                        sizeof(key->enc_key_id.keyid));
1204
1205         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1206                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1207                        sizeof(key->enc_tp.src));
1208
1209         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1210                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1211                        sizeof(key->enc_tp.dst));
1212
1213         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1214
1215         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1216                 ret = fl_set_enc_opt(tb, key, mask, extack);
1217                 if (ret)
1218                         return ret;
1219         }
1220
1221         if (tb[TCA_FLOWER_KEY_FLAGS])
1222                 ret = fl_set_key_flags(tb, &key->control.flags, &mask->control.flags);
1223
1224         return ret;
1225 }
1226
1227 static void fl_mask_copy(struct fl_flow_mask *dst,
1228                          struct fl_flow_mask *src)
1229 {
1230         const void *psrc = fl_key_get_start(&src->key, src);
1231         void *pdst = fl_key_get_start(&dst->key, src);
1232
1233         memcpy(pdst, psrc, fl_mask_range(src));
1234         dst->range = src->range;
1235 }
1236
1237 static const struct rhashtable_params fl_ht_params = {
1238         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1239         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1240         .automatic_shrinking = true,
1241 };
1242
1243 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1244 {
1245         mask->filter_ht_params = fl_ht_params;
1246         mask->filter_ht_params.key_len = fl_mask_range(mask);
1247         mask->filter_ht_params.key_offset += mask->range.start;
1248
1249         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1250 }
1251
1252 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1253 #define FL_KEY_MEMBER_SIZE(member) FIELD_SIZEOF(struct fl_flow_key, member)
1254
1255 #define FL_KEY_IS_MASKED(mask, member)                                          \
1256         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1257                    0, FL_KEY_MEMBER_SIZE(member))                               \
1258
1259 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1260         do {                                                                    \
1261                 keys[cnt].key_id = id;                                          \
1262                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1263                 cnt++;                                                          \
1264         } while(0);
1265
1266 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1267         do {                                                                    \
1268                 if (FL_KEY_IS_MASKED(mask, member))                             \
1269                         FL_KEY_SET(keys, cnt, id, member);                      \
1270         } while(0);
1271
1272 static void fl_init_dissector(struct flow_dissector *dissector,
1273                               struct fl_flow_key *mask)
1274 {
1275         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1276         size_t cnt = 0;
1277
1278         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1279         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1280         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1281                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1282         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1283                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1284         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1285                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1286         if (FL_KEY_IS_MASKED(mask, tp) ||
1287             FL_KEY_IS_MASKED(mask, tp_min) || FL_KEY_IS_MASKED(mask, tp_max))
1288                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_PORTS, tp);
1289         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1290                              FLOW_DISSECTOR_KEY_IP, ip);
1291         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1292                              FLOW_DISSECTOR_KEY_TCP, tcp);
1293         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1294                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1295         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1296                              FLOW_DISSECTOR_KEY_ARP, arp);
1297         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1298                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1299         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1300                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1301         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1302                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1303         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1304                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1305         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1306                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1307         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1308                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1309         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1310             FL_KEY_IS_MASKED(mask, enc_ipv6))
1311                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1312                            enc_control);
1313         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1314                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1315         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1316                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1317         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1318                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1319
1320         skb_flow_dissector_init(dissector, keys, cnt);
1321 }
1322
1323 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1324                                                struct fl_flow_mask *mask)
1325 {
1326         struct fl_flow_mask *newmask;
1327         int err;
1328
1329         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1330         if (!newmask)
1331                 return ERR_PTR(-ENOMEM);
1332
1333         fl_mask_copy(newmask, mask);
1334
1335         if ((newmask->key.tp_min.dst && newmask->key.tp_max.dst) ||
1336             (newmask->key.tp_min.src && newmask->key.tp_max.src))
1337                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1338
1339         err = fl_init_mask_hashtable(newmask);
1340         if (err)
1341                 goto errout_free;
1342
1343         fl_init_dissector(&newmask->dissector, &newmask->key);
1344
1345         INIT_LIST_HEAD_RCU(&newmask->filters);
1346
1347         refcount_set(&newmask->refcnt, 1);
1348         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1349                                       &newmask->ht_node, mask_ht_params);
1350         if (err)
1351                 goto errout_destroy;
1352
1353         /* Wait until any potential concurrent users of mask are finished */
1354         synchronize_rcu();
1355
1356         spin_lock(&head->masks_lock);
1357         list_add_tail_rcu(&newmask->list, &head->masks);
1358         spin_unlock(&head->masks_lock);
1359
1360         return newmask;
1361
1362 errout_destroy:
1363         rhashtable_destroy(&newmask->ht);
1364 errout_free:
1365         kfree(newmask);
1366
1367         return ERR_PTR(err);
1368 }
1369
1370 static int fl_check_assign_mask(struct cls_fl_head *head,
1371                                 struct cls_fl_filter *fnew,
1372                                 struct cls_fl_filter *fold,
1373                                 struct fl_flow_mask *mask)
1374 {
1375         struct fl_flow_mask *newmask;
1376         int ret = 0;
1377
1378         rcu_read_lock();
1379
1380         /* Insert mask as temporary node to prevent concurrent creation of mask
1381          * with same key. Any concurrent lookups with same key will return
1382          * -EAGAIN because mask's refcnt is zero. It is safe to insert
1383          * stack-allocated 'mask' to masks hash table because we call
1384          * synchronize_rcu() before returning from this function (either in case
1385          * of error or after replacing it with heap-allocated mask in
1386          * fl_create_new_mask()).
1387          */
1388         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1389                                                        &mask->ht_node,
1390                                                        mask_ht_params);
1391         if (!fnew->mask) {
1392                 rcu_read_unlock();
1393
1394                 if (fold) {
1395                         ret = -EINVAL;
1396                         goto errout_cleanup;
1397                 }
1398
1399                 newmask = fl_create_new_mask(head, mask);
1400                 if (IS_ERR(newmask)) {
1401                         ret = PTR_ERR(newmask);
1402                         goto errout_cleanup;
1403                 }
1404
1405                 fnew->mask = newmask;
1406                 return 0;
1407         } else if (IS_ERR(fnew->mask)) {
1408                 ret = PTR_ERR(fnew->mask);
1409         } else if (fold && fold->mask != fnew->mask) {
1410                 ret = -EINVAL;
1411         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1412                 /* Mask was deleted concurrently, try again */
1413                 ret = -EAGAIN;
1414         }
1415         rcu_read_unlock();
1416         return ret;
1417
1418 errout_cleanup:
1419         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1420                                mask_ht_params);
1421         /* Wait until any potential concurrent users of mask are finished */
1422         synchronize_rcu();
1423         return ret;
1424 }
1425
1426 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1427                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1428                         unsigned long base, struct nlattr **tb,
1429                         struct nlattr *est, bool ovr,
1430                         struct fl_flow_tmplt *tmplt, bool rtnl_held,
1431                         struct netlink_ext_ack *extack)
1432 {
1433         int err;
1434
1435         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, rtnl_held,
1436                                 extack);
1437         if (err < 0)
1438                 return err;
1439
1440         if (tb[TCA_FLOWER_CLASSID]) {
1441                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1442                 if (!rtnl_held)
1443                         rtnl_lock();
1444                 tcf_bind_filter(tp, &f->res, base);
1445                 if (!rtnl_held)
1446                         rtnl_unlock();
1447         }
1448
1449         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1450         if (err)
1451                 return err;
1452
1453         fl_mask_update_range(mask);
1454         fl_set_masked_key(&f->mkey, &f->key, mask);
1455
1456         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1457                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1458                 return -EINVAL;
1459         }
1460
1461         return 0;
1462 }
1463
1464 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
1465                                struct cls_fl_filter *fold,
1466                                bool *in_ht)
1467 {
1468         struct fl_flow_mask *mask = fnew->mask;
1469         int err;
1470
1471         err = rhashtable_lookup_insert_fast(&mask->ht,
1472                                             &fnew->ht_node,
1473                                             mask->filter_ht_params);
1474         if (err) {
1475                 *in_ht = false;
1476                 /* It is okay if filter with same key exists when
1477                  * overwriting.
1478                  */
1479                 return fold && err == -EEXIST ? 0 : err;
1480         }
1481
1482         *in_ht = true;
1483         return 0;
1484 }
1485
1486 static int fl_change(struct net *net, struct sk_buff *in_skb,
1487                      struct tcf_proto *tp, unsigned long base,
1488                      u32 handle, struct nlattr **tca,
1489                      void **arg, bool ovr, bool rtnl_held,
1490                      struct netlink_ext_ack *extack)
1491 {
1492         struct cls_fl_head *head = fl_head_dereference(tp);
1493         struct cls_fl_filter *fold = *arg;
1494         struct cls_fl_filter *fnew;
1495         struct fl_flow_mask *mask;
1496         struct nlattr **tb;
1497         bool in_ht;
1498         int err;
1499
1500         if (!tca[TCA_OPTIONS]) {
1501                 err = -EINVAL;
1502                 goto errout_fold;
1503         }
1504
1505         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
1506         if (!mask) {
1507                 err = -ENOBUFS;
1508                 goto errout_fold;
1509         }
1510
1511         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1512         if (!tb) {
1513                 err = -ENOBUFS;
1514                 goto errout_mask_alloc;
1515         }
1516
1517         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1518                                           tca[TCA_OPTIONS], fl_policy, NULL);
1519         if (err < 0)
1520                 goto errout_tb;
1521
1522         if (fold && handle && fold->handle != handle) {
1523                 err = -EINVAL;
1524                 goto errout_tb;
1525         }
1526
1527         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
1528         if (!fnew) {
1529                 err = -ENOBUFS;
1530                 goto errout_tb;
1531         }
1532         INIT_LIST_HEAD(&fnew->hw_list);
1533         refcount_set(&fnew->refcnt, 1);
1534
1535         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
1536         if (err < 0)
1537                 goto errout;
1538
1539         if (tb[TCA_FLOWER_FLAGS]) {
1540                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
1541
1542                 if (!tc_flags_valid(fnew->flags)) {
1543                         err = -EINVAL;
1544                         goto errout;
1545                 }
1546         }
1547
1548         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE], ovr,
1549                            tp->chain->tmplt_priv, rtnl_held, extack);
1550         if (err)
1551                 goto errout;
1552
1553         err = fl_check_assign_mask(head, fnew, fold, mask);
1554         if (err)
1555                 goto errout;
1556
1557         err = fl_ht_insert_unique(fnew, fold, &in_ht);
1558         if (err)
1559                 goto errout_mask;
1560
1561         if (!tc_skip_hw(fnew->flags)) {
1562                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
1563                 if (err)
1564                         goto errout_ht;
1565         }
1566
1567         if (!tc_in_hw(fnew->flags))
1568                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1569
1570         spin_lock(&tp->lock);
1571
1572         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
1573          * proto again or create new one, if necessary.
1574          */
1575         if (tp->deleting) {
1576                 err = -EAGAIN;
1577                 goto errout_hw;
1578         }
1579
1580         if (fold) {
1581                 /* Fold filter was deleted concurrently. Retry lookup. */
1582                 if (fold->deleted) {
1583                         err = -EAGAIN;
1584                         goto errout_hw;
1585                 }
1586
1587                 fnew->handle = handle;
1588
1589                 if (!in_ht) {
1590                         struct rhashtable_params params =
1591                                 fnew->mask->filter_ht_params;
1592
1593                         err = rhashtable_insert_fast(&fnew->mask->ht,
1594                                                      &fnew->ht_node,
1595                                                      params);
1596                         if (err)
1597                                 goto errout_hw;
1598                         in_ht = true;
1599                 }
1600
1601                 refcount_inc(&fnew->refcnt);
1602                 rhashtable_remove_fast(&fold->mask->ht,
1603                                        &fold->ht_node,
1604                                        fold->mask->filter_ht_params);
1605                 idr_replace(&head->handle_idr, fnew, fnew->handle);
1606                 list_replace_rcu(&fold->list, &fnew->list);
1607                 fold->deleted = true;
1608
1609                 spin_unlock(&tp->lock);
1610
1611                 fl_mask_put(head, fold->mask);
1612                 if (!tc_skip_hw(fold->flags))
1613                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
1614                 tcf_unbind_filter(tp, &fold->res);
1615                 /* Caller holds reference to fold, so refcnt is always > 0
1616                  * after this.
1617                  */
1618                 refcount_dec(&fold->refcnt);
1619                 __fl_put(fold);
1620         } else {
1621                 if (handle) {
1622                         /* user specifies a handle and it doesn't exist */
1623                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1624                                             handle, GFP_ATOMIC);
1625
1626                         /* Filter with specified handle was concurrently
1627                          * inserted after initial check in cls_api. This is not
1628                          * necessarily an error if NLM_F_EXCL is not set in
1629                          * message flags. Returning EAGAIN will cause cls_api to
1630                          * try to update concurrently inserted rule.
1631                          */
1632                         if (err == -ENOSPC)
1633                                 err = -EAGAIN;
1634                 } else {
1635                         handle = 1;
1636                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1637                                             INT_MAX, GFP_ATOMIC);
1638                 }
1639                 if (err)
1640                         goto errout_hw;
1641
1642                 refcount_inc(&fnew->refcnt);
1643                 fnew->handle = handle;
1644                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
1645                 spin_unlock(&tp->lock);
1646         }
1647
1648         *arg = fnew;
1649
1650         kfree(tb);
1651         kfree(mask);
1652         return 0;
1653
1654 errout_ht:
1655         spin_lock(&tp->lock);
1656 errout_hw:
1657         fnew->deleted = true;
1658         spin_unlock(&tp->lock);
1659         if (!tc_skip_hw(fnew->flags))
1660                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
1661         if (in_ht)
1662                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
1663                                        fnew->mask->filter_ht_params);
1664 errout_mask:
1665         fl_mask_put(head, fnew->mask);
1666 errout:
1667         __fl_put(fnew);
1668 errout_tb:
1669         kfree(tb);
1670 errout_mask_alloc:
1671         kfree(mask);
1672 errout_fold:
1673         if (fold)
1674                 __fl_put(fold);
1675         return err;
1676 }
1677
1678 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
1679                      bool rtnl_held, struct netlink_ext_ack *extack)
1680 {
1681         struct cls_fl_head *head = fl_head_dereference(tp);
1682         struct cls_fl_filter *f = arg;
1683         bool last_on_mask;
1684         int err = 0;
1685
1686         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
1687         *last = list_empty(&head->masks);
1688         __fl_put(f);
1689
1690         return err;
1691 }
1692
1693 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
1694                     bool rtnl_held)
1695 {
1696         struct cls_fl_filter *f;
1697
1698         arg->count = arg->skip;
1699
1700         while ((f = fl_get_next_filter(tp, &arg->cookie)) != NULL) {
1701                 if (arg->fn(tp, f, arg) < 0) {
1702                         __fl_put(f);
1703                         arg->stop = 1;
1704                         break;
1705                 }
1706                 __fl_put(f);
1707                 arg->cookie++;
1708                 arg->count++;
1709         }
1710 }
1711
1712 static struct cls_fl_filter *
1713 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
1714 {
1715         struct cls_fl_head *head = fl_head_dereference(tp);
1716
1717         spin_lock(&tp->lock);
1718         if (list_empty(&head->hw_filters)) {
1719                 spin_unlock(&tp->lock);
1720                 return NULL;
1721         }
1722
1723         if (!f)
1724                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
1725                                hw_list);
1726         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
1727                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
1728                         spin_unlock(&tp->lock);
1729                         return f;
1730                 }
1731         }
1732
1733         spin_unlock(&tp->lock);
1734         return NULL;
1735 }
1736
1737 static int fl_reoffload(struct tcf_proto *tp, bool add, tc_setup_cb_t *cb,
1738                         void *cb_priv, struct netlink_ext_ack *extack)
1739 {
1740         struct tc_cls_flower_offload cls_flower = {};
1741         struct tcf_block *block = tp->chain->block;
1742         struct cls_fl_filter *f = NULL;
1743         int err;
1744
1745         /* hw_filters list can only be changed by hw offload functions after
1746          * obtaining rtnl lock. Make sure it is not changed while reoffload is
1747          * iterating it.
1748          */
1749         ASSERT_RTNL();
1750
1751         while ((f = fl_get_next_hw_filter(tp, f, add))) {
1752                 cls_flower.rule =
1753                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
1754                 if (!cls_flower.rule) {
1755                         __fl_put(f);
1756                         return -ENOMEM;
1757                 }
1758
1759                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
1760                                            extack);
1761                 cls_flower.command = add ?
1762                         TC_CLSFLOWER_REPLACE : TC_CLSFLOWER_DESTROY;
1763                 cls_flower.cookie = (unsigned long)f;
1764                 cls_flower.rule->match.dissector = &f->mask->dissector;
1765                 cls_flower.rule->match.mask = &f->mask->key;
1766                 cls_flower.rule->match.key = &f->mkey;
1767
1768                 err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
1769                 if (err) {
1770                         kfree(cls_flower.rule);
1771                         if (tc_skip_sw(f->flags)) {
1772                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
1773                                 __fl_put(f);
1774                                 return err;
1775                         }
1776                         goto next_flow;
1777                 }
1778
1779                 cls_flower.classid = f->res.classid;
1780
1781                 err = cb(TC_SETUP_CLSFLOWER, &cls_flower, cb_priv);
1782                 kfree(cls_flower.rule);
1783
1784                 if (err) {
1785                         if (add && tc_skip_sw(f->flags)) {
1786                                 __fl_put(f);
1787                                 return err;
1788                         }
1789                         goto next_flow;
1790                 }
1791
1792                 spin_lock(&tp->lock);
1793                 tc_cls_offload_cnt_update(block, &f->in_hw_count, &f->flags,
1794                                           add);
1795                 spin_unlock(&tp->lock);
1796 next_flow:
1797                 __fl_put(f);
1798         }
1799
1800         return 0;
1801 }
1802
1803 static int fl_hw_create_tmplt(struct tcf_chain *chain,
1804                               struct fl_flow_tmplt *tmplt)
1805 {
1806         struct tc_cls_flower_offload cls_flower = {};
1807         struct tcf_block *block = chain->block;
1808
1809         cls_flower.rule = flow_rule_alloc(0);
1810         if (!cls_flower.rule)
1811                 return -ENOMEM;
1812
1813         cls_flower.common.chain_index = chain->index;
1814         cls_flower.command = TC_CLSFLOWER_TMPLT_CREATE;
1815         cls_flower.cookie = (unsigned long) tmplt;
1816         cls_flower.rule->match.dissector = &tmplt->dissector;
1817         cls_flower.rule->match.mask = &tmplt->mask;
1818         cls_flower.rule->match.key = &tmplt->dummy_key;
1819
1820         /* We don't care if driver (any of them) fails to handle this
1821          * call. It serves just as a hint for it.
1822          */
1823         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
1824         kfree(cls_flower.rule);
1825
1826         return 0;
1827 }
1828
1829 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
1830                                 struct fl_flow_tmplt *tmplt)
1831 {
1832         struct tc_cls_flower_offload cls_flower = {};
1833         struct tcf_block *block = chain->block;
1834
1835         cls_flower.common.chain_index = chain->index;
1836         cls_flower.command = TC_CLSFLOWER_TMPLT_DESTROY;
1837         cls_flower.cookie = (unsigned long) tmplt;
1838
1839         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
1840 }
1841
1842 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
1843                              struct nlattr **tca,
1844                              struct netlink_ext_ack *extack)
1845 {
1846         struct fl_flow_tmplt *tmplt;
1847         struct nlattr **tb;
1848         int err;
1849
1850         if (!tca[TCA_OPTIONS])
1851                 return ERR_PTR(-EINVAL);
1852
1853         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1854         if (!tb)
1855                 return ERR_PTR(-ENOBUFS);
1856         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1857                                           tca[TCA_OPTIONS], fl_policy, NULL);
1858         if (err)
1859                 goto errout_tb;
1860
1861         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
1862         if (!tmplt) {
1863                 err = -ENOMEM;
1864                 goto errout_tb;
1865         }
1866         tmplt->chain = chain;
1867         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
1868         if (err)
1869                 goto errout_tmplt;
1870
1871         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
1872
1873         err = fl_hw_create_tmplt(chain, tmplt);
1874         if (err)
1875                 goto errout_tmplt;
1876
1877         kfree(tb);
1878         return tmplt;
1879
1880 errout_tmplt:
1881         kfree(tmplt);
1882 errout_tb:
1883         kfree(tb);
1884         return ERR_PTR(err);
1885 }
1886
1887 static void fl_tmplt_destroy(void *tmplt_priv)
1888 {
1889         struct fl_flow_tmplt *tmplt = tmplt_priv;
1890
1891         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
1892         kfree(tmplt);
1893 }
1894
1895 static int fl_dump_key_val(struct sk_buff *skb,
1896                            void *val, int val_type,
1897                            void *mask, int mask_type, int len)
1898 {
1899         int err;
1900
1901         if (!memchr_inv(mask, 0, len))
1902                 return 0;
1903         err = nla_put(skb, val_type, len, val);
1904         if (err)
1905                 return err;
1906         if (mask_type != TCA_FLOWER_UNSPEC) {
1907                 err = nla_put(skb, mask_type, len, mask);
1908                 if (err)
1909                         return err;
1910         }
1911         return 0;
1912 }
1913
1914 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
1915                                   struct fl_flow_key *mask)
1916 {
1917         if (fl_dump_key_val(skb, &key->tp_min.dst, TCA_FLOWER_KEY_PORT_DST_MIN,
1918                             &mask->tp_min.dst, TCA_FLOWER_UNSPEC,
1919                             sizeof(key->tp_min.dst)) ||
1920             fl_dump_key_val(skb, &key->tp_max.dst, TCA_FLOWER_KEY_PORT_DST_MAX,
1921                             &mask->tp_max.dst, TCA_FLOWER_UNSPEC,
1922                             sizeof(key->tp_max.dst)) ||
1923             fl_dump_key_val(skb, &key->tp_min.src, TCA_FLOWER_KEY_PORT_SRC_MIN,
1924                             &mask->tp_min.src, TCA_FLOWER_UNSPEC,
1925                             sizeof(key->tp_min.src)) ||
1926             fl_dump_key_val(skb, &key->tp_max.src, TCA_FLOWER_KEY_PORT_SRC_MAX,
1927                             &mask->tp_max.src, TCA_FLOWER_UNSPEC,
1928                             sizeof(key->tp_max.src)))
1929                 return -1;
1930
1931         return 0;
1932 }
1933
1934 static int fl_dump_key_mpls(struct sk_buff *skb,
1935                             struct flow_dissector_key_mpls *mpls_key,
1936                             struct flow_dissector_key_mpls *mpls_mask)
1937 {
1938         int err;
1939
1940         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
1941                 return 0;
1942         if (mpls_mask->mpls_ttl) {
1943                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
1944                                  mpls_key->mpls_ttl);
1945                 if (err)
1946                         return err;
1947         }
1948         if (mpls_mask->mpls_tc) {
1949                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
1950                                  mpls_key->mpls_tc);
1951                 if (err)
1952                         return err;
1953         }
1954         if (mpls_mask->mpls_label) {
1955                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
1956                                   mpls_key->mpls_label);
1957                 if (err)
1958                         return err;
1959         }
1960         if (mpls_mask->mpls_bos) {
1961                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
1962                                  mpls_key->mpls_bos);
1963                 if (err)
1964                         return err;
1965         }
1966         return 0;
1967 }
1968
1969 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
1970                           struct flow_dissector_key_ip *key,
1971                           struct flow_dissector_key_ip *mask)
1972 {
1973         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1974         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1975         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1976         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1977
1978         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
1979             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
1980                 return -1;
1981
1982         return 0;
1983 }
1984
1985 static int fl_dump_key_vlan(struct sk_buff *skb,
1986                             int vlan_id_key, int vlan_prio_key,
1987                             struct flow_dissector_key_vlan *vlan_key,
1988                             struct flow_dissector_key_vlan *vlan_mask)
1989 {
1990         int err;
1991
1992         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
1993                 return 0;
1994         if (vlan_mask->vlan_id) {
1995                 err = nla_put_u16(skb, vlan_id_key,
1996                                   vlan_key->vlan_id);
1997                 if (err)
1998                         return err;
1999         }
2000         if (vlan_mask->vlan_priority) {
2001                 err = nla_put_u8(skb, vlan_prio_key,
2002                                  vlan_key->vlan_priority);
2003                 if (err)
2004                         return err;
2005         }
2006         return 0;
2007 }
2008
2009 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2010                             u32 *flower_key, u32 *flower_mask,
2011                             u32 flower_flag_bit, u32 dissector_flag_bit)
2012 {
2013         if (dissector_mask & dissector_flag_bit) {
2014                 *flower_mask |= flower_flag_bit;
2015                 if (dissector_key & dissector_flag_bit)
2016                         *flower_key |= flower_flag_bit;
2017         }
2018 }
2019
2020 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2021 {
2022         u32 key, mask;
2023         __be32 _key, _mask;
2024         int err;
2025
2026         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2027                 return 0;
2028
2029         key = 0;
2030         mask = 0;
2031
2032         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2033                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2034         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2035                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2036                         FLOW_DIS_FIRST_FRAG);
2037
2038         _key = cpu_to_be32(key);
2039         _mask = cpu_to_be32(mask);
2040
2041         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2042         if (err)
2043                 return err;
2044
2045         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2046 }
2047
2048 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2049                                   struct flow_dissector_key_enc_opts *enc_opts)
2050 {
2051         struct geneve_opt *opt;
2052         struct nlattr *nest;
2053         int opt_off = 0;
2054
2055         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2056         if (!nest)
2057                 goto nla_put_failure;
2058
2059         while (enc_opts->len > opt_off) {
2060                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2061
2062                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2063                                  opt->opt_class))
2064                         goto nla_put_failure;
2065                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2066                                opt->type))
2067                         goto nla_put_failure;
2068                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2069                             opt->length * 4, opt->opt_data))
2070                         goto nla_put_failure;
2071
2072                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2073         }
2074         nla_nest_end(skb, nest);
2075         return 0;
2076
2077 nla_put_failure:
2078         nla_nest_cancel(skb, nest);
2079         return -EMSGSIZE;
2080 }
2081
2082 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2083                                struct flow_dissector_key_enc_opts *enc_opts)
2084 {
2085         struct nlattr *nest;
2086         int err;
2087
2088         if (!enc_opts->len)
2089                 return 0;
2090
2091         nest = nla_nest_start_noflag(skb, enc_opt_type);
2092         if (!nest)
2093                 goto nla_put_failure;
2094
2095         switch (enc_opts->dst_opt_type) {
2096         case TUNNEL_GENEVE_OPT:
2097                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2098                 if (err)
2099                         goto nla_put_failure;
2100                 break;
2101         default:
2102                 goto nla_put_failure;
2103         }
2104         nla_nest_end(skb, nest);
2105         return 0;
2106
2107 nla_put_failure:
2108         nla_nest_cancel(skb, nest);
2109         return -EMSGSIZE;
2110 }
2111
2112 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2113                                struct flow_dissector_key_enc_opts *key_opts,
2114                                struct flow_dissector_key_enc_opts *msk_opts)
2115 {
2116         int err;
2117
2118         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2119         if (err)
2120                 return err;
2121
2122         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2123 }
2124
2125 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2126                        struct fl_flow_key *key, struct fl_flow_key *mask)
2127 {
2128         if (mask->indev_ifindex) {
2129                 struct net_device *dev;
2130
2131                 dev = __dev_get_by_index(net, key->indev_ifindex);
2132                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2133                         goto nla_put_failure;
2134         }
2135
2136         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2137                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2138                             sizeof(key->eth.dst)) ||
2139             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2140                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2141                             sizeof(key->eth.src)) ||
2142             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2143                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2144                             sizeof(key->basic.n_proto)))
2145                 goto nla_put_failure;
2146
2147         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2148                 goto nla_put_failure;
2149
2150         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2151                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2152                 goto nla_put_failure;
2153
2154         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2155                              TCA_FLOWER_KEY_CVLAN_PRIO,
2156                              &key->cvlan, &mask->cvlan) ||
2157             (mask->cvlan.vlan_tpid &&
2158              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2159                           key->cvlan.vlan_tpid)))
2160                 goto nla_put_failure;
2161
2162         if (mask->basic.n_proto) {
2163                 if (mask->cvlan.vlan_tpid) {
2164                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
2165                                          key->basic.n_proto))
2166                                 goto nla_put_failure;
2167                 } else if (mask->vlan.vlan_tpid) {
2168                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2169                                          key->basic.n_proto))
2170                                 goto nla_put_failure;
2171                 }
2172         }
2173
2174         if ((key->basic.n_proto == htons(ETH_P_IP) ||
2175              key->basic.n_proto == htons(ETH_P_IPV6)) &&
2176             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
2177                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
2178                             sizeof(key->basic.ip_proto)) ||
2179             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
2180                 goto nla_put_failure;
2181
2182         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2183             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
2184                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
2185                              sizeof(key->ipv4.src)) ||
2186              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
2187                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
2188                              sizeof(key->ipv4.dst))))
2189                 goto nla_put_failure;
2190         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2191                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
2192                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
2193                                   sizeof(key->ipv6.src)) ||
2194                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
2195                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
2196                                   sizeof(key->ipv6.dst))))
2197                 goto nla_put_failure;
2198
2199         if (key->basic.ip_proto == IPPROTO_TCP &&
2200             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
2201                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
2202                              sizeof(key->tp.src)) ||
2203              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
2204                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
2205                              sizeof(key->tp.dst)) ||
2206              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
2207                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
2208                              sizeof(key->tcp.flags))))
2209                 goto nla_put_failure;
2210         else if (key->basic.ip_proto == IPPROTO_UDP &&
2211                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
2212                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
2213                                   sizeof(key->tp.src)) ||
2214                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
2215                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
2216                                   sizeof(key->tp.dst))))
2217                 goto nla_put_failure;
2218         else if (key->basic.ip_proto == IPPROTO_SCTP &&
2219                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
2220                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
2221                                   sizeof(key->tp.src)) ||
2222                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
2223                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
2224                                   sizeof(key->tp.dst))))
2225                 goto nla_put_failure;
2226         else if (key->basic.n_proto == htons(ETH_P_IP) &&
2227                  key->basic.ip_proto == IPPROTO_ICMP &&
2228                  (fl_dump_key_val(skb, &key->icmp.type,
2229                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
2230                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
2231                                   sizeof(key->icmp.type)) ||
2232                   fl_dump_key_val(skb, &key->icmp.code,
2233                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
2234                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
2235                                   sizeof(key->icmp.code))))
2236                 goto nla_put_failure;
2237         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
2238                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
2239                  (fl_dump_key_val(skb, &key->icmp.type,
2240                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
2241                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
2242                                   sizeof(key->icmp.type)) ||
2243                   fl_dump_key_val(skb, &key->icmp.code,
2244                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
2245                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
2246                                   sizeof(key->icmp.code))))
2247                 goto nla_put_failure;
2248         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
2249                   key->basic.n_proto == htons(ETH_P_RARP)) &&
2250                  (fl_dump_key_val(skb, &key->arp.sip,
2251                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
2252                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
2253                                   sizeof(key->arp.sip)) ||
2254                   fl_dump_key_val(skb, &key->arp.tip,
2255                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
2256                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
2257                                   sizeof(key->arp.tip)) ||
2258                   fl_dump_key_val(skb, &key->arp.op,
2259                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
2260                                   TCA_FLOWER_KEY_ARP_OP_MASK,
2261                                   sizeof(key->arp.op)) ||
2262                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2263                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2264                                   sizeof(key->arp.sha)) ||
2265                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2266                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2267                                   sizeof(key->arp.tha))))
2268                 goto nla_put_failure;
2269
2270         if ((key->basic.ip_proto == IPPROTO_TCP ||
2271              key->basic.ip_proto == IPPROTO_UDP ||
2272              key->basic.ip_proto == IPPROTO_SCTP) &&
2273              fl_dump_key_port_range(skb, key, mask))
2274                 goto nla_put_failure;
2275
2276         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2277             (fl_dump_key_val(skb, &key->enc_ipv4.src,
2278                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
2279                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
2280                             sizeof(key->enc_ipv4.src)) ||
2281              fl_dump_key_val(skb, &key->enc_ipv4.dst,
2282                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
2283                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
2284                              sizeof(key->enc_ipv4.dst))))
2285                 goto nla_put_failure;
2286         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2287                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
2288                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
2289                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
2290                             sizeof(key->enc_ipv6.src)) ||
2291                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
2292                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
2293                                  &mask->enc_ipv6.dst,
2294                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
2295                             sizeof(key->enc_ipv6.dst))))
2296                 goto nla_put_failure;
2297
2298         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
2299                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
2300                             sizeof(key->enc_key_id)) ||
2301             fl_dump_key_val(skb, &key->enc_tp.src,
2302                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
2303                             &mask->enc_tp.src,
2304                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
2305                             sizeof(key->enc_tp.src)) ||
2306             fl_dump_key_val(skb, &key->enc_tp.dst,
2307                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
2308                             &mask->enc_tp.dst,
2309                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
2310                             sizeof(key->enc_tp.dst)) ||
2311             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
2312             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
2313                 goto nla_put_failure;
2314
2315         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
2316                 goto nla_put_failure;
2317
2318         return 0;
2319
2320 nla_put_failure:
2321         return -EMSGSIZE;
2322 }
2323
2324 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
2325                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
2326 {
2327         struct cls_fl_filter *f = fh;
2328         struct nlattr *nest;
2329         struct fl_flow_key *key, *mask;
2330         bool skip_hw;
2331
2332         if (!f)
2333                 return skb->len;
2334
2335         t->tcm_handle = f->handle;
2336
2337         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2338         if (!nest)
2339                 goto nla_put_failure;
2340
2341         spin_lock(&tp->lock);
2342
2343         if (f->res.classid &&
2344             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
2345                 goto nla_put_failure_locked;
2346
2347         key = &f->key;
2348         mask = &f->mask->key;
2349         skip_hw = tc_skip_hw(f->flags);
2350
2351         if (fl_dump_key(skb, net, key, mask))
2352                 goto nla_put_failure_locked;
2353
2354         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
2355                 goto nla_put_failure_locked;
2356
2357         spin_unlock(&tp->lock);
2358
2359         if (!skip_hw)
2360                 fl_hw_update_stats(tp, f, rtnl_held);
2361
2362         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
2363                 goto nla_put_failure;
2364
2365         if (tcf_exts_dump(skb, &f->exts))
2366                 goto nla_put_failure;
2367
2368         nla_nest_end(skb, nest);
2369
2370         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
2371                 goto nla_put_failure;
2372
2373         return skb->len;
2374
2375 nla_put_failure_locked:
2376         spin_unlock(&tp->lock);
2377 nla_put_failure:
2378         nla_nest_cancel(skb, nest);
2379         return -1;
2380 }
2381
2382 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
2383 {
2384         struct fl_flow_tmplt *tmplt = tmplt_priv;
2385         struct fl_flow_key *key, *mask;
2386         struct nlattr *nest;
2387
2388         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2389         if (!nest)
2390                 goto nla_put_failure;
2391
2392         key = &tmplt->dummy_key;
2393         mask = &tmplt->mask;
2394
2395         if (fl_dump_key(skb, net, key, mask))
2396                 goto nla_put_failure;
2397
2398         nla_nest_end(skb, nest);
2399
2400         return skb->len;
2401
2402 nla_put_failure:
2403         nla_nest_cancel(skb, nest);
2404         return -EMSGSIZE;
2405 }
2406
2407 static void fl_bind_class(void *fh, u32 classid, unsigned long cl)
2408 {
2409         struct cls_fl_filter *f = fh;
2410
2411         if (f && f->res.classid == classid)
2412                 f->res.class = cl;
2413 }
2414
2415 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
2416         .kind           = "flower",
2417         .classify       = fl_classify,
2418         .init           = fl_init,
2419         .destroy        = fl_destroy,
2420         .get            = fl_get,
2421         .put            = fl_put,
2422         .change         = fl_change,
2423         .delete         = fl_delete,
2424         .walk           = fl_walk,
2425         .reoffload      = fl_reoffload,
2426         .dump           = fl_dump,
2427         .bind_class     = fl_bind_class,
2428         .tmplt_create   = fl_tmplt_create,
2429         .tmplt_destroy  = fl_tmplt_destroy,
2430         .tmplt_dump     = fl_tmplt_dump,
2431         .owner          = THIS_MODULE,
2432         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
2433 };
2434
2435 static int __init cls_fl_init(void)
2436 {
2437         return register_tcf_proto_ops(&cls_fl_ops);
2438 }
2439
2440 static void __exit cls_fl_exit(void)
2441 {
2442         unregister_tcf_proto_ops(&cls_fl_ops);
2443 }
2444
2445 module_init(cls_fl_init);
2446 module_exit(cls_fl_exit);
2447
2448 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
2449 MODULE_DESCRIPTION("Flower classifier");
2450 MODULE_LICENSE("GPL v2");