net/sched/cls_flower.c

   1 /*
   2  * net/sched/cls_flower.c               Flower classifier
   3  *
   4  * Copyright (c) 2015 Jiri Pirko <[email protected]>
   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
  18 #include <linux/if_ether.h>
  19 #include <linux/in6.h>
  20 #include <linux/ip.h>
  21 #include <linux/mpls.h>
  22
  23 #include <net/sch_generic.h>
  24 #include <net/pkt_cls.h>
  25 #include <net/ip.h>
  26 #include <net/flow_dissector.h>
  27 #include <net/geneve.h>
  28
  29 #include <net/dst.h>
  30 #include <net/dst_metadata.h>
  31
  32 struct fl_flow_key {
  33         int     indev_ifindex;
  34         struct flow_dissector_key_control control;
  35         struct flow_dissector_key_control enc_control;
  36         struct flow_dissector_key_basic basic;
  37         struct flow_dissector_key_eth_addrs eth;
  38         struct flow_dissector_key_vlan vlan;
  39         struct flow_dissector_key_vlan cvlan;
  40         union {
  41                 struct flow_dissector_key_ipv4_addrs ipv4;
  42                 struct flow_dissector_key_ipv6_addrs ipv6;
  43         };
  44         struct flow_dissector_key_ports tp;
  45         struct flow_dissector_key_icmp icmp;
  46         struct flow_dissector_key_arp arp;
  47         struct flow_dissector_key_keyid enc_key_id;
  48         union {
  49                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  50                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  51         };
  52         struct flow_dissector_key_ports enc_tp;
  53         struct flow_dissector_key_mpls mpls;
  54         struct flow_dissector_key_tcp tcp;
  55         struct flow_dissector_key_ip ip;
  56         struct flow_dissector_key_ip enc_ip;
  57         struct flow_dissector_key_enc_opts enc_opts;
  58 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  59
  60 struct fl_flow_mask_range {
  61         unsigned short int start;
  62         unsigned short int end;
  63 };
  64
  65 struct fl_flow_mask {
  66         struct fl_flow_key key;
  67         struct fl_flow_mask_range range;
  68         struct rhash_head ht_node;
  69         struct rhashtable ht;
  70         struct rhashtable_params filter_ht_params;
  71         struct flow_dissector dissector;
  72         struct list_head filters;
  73         struct rcu_work rwork;
  74         struct list_head list;
  75 };
  76
  77 struct fl_flow_tmplt {
  78         struct fl_flow_key dummy_key;
  79         struct fl_flow_key mask;
  80         struct flow_dissector dissector;
  81         struct tcf_chain *chain;
  82 };
  83
  84 struct cls_fl_head {
  85         struct rhashtable ht;
  86         struct list_head masks;
  87         struct rcu_work rwork;
  88         struct idr handle_idr;
  89 };
  90
  91 struct cls_fl_filter {
  92         struct fl_flow_mask *mask;
  93         struct rhash_head ht_node;
  94         struct fl_flow_key mkey;
  95         struct tcf_exts exts;
  96         struct tcf_result res;
  97         struct fl_flow_key key;
  98         struct list_head list;
  99         u32 handle;
 100         u32 flags;
 101         unsigned int in_hw_count;
 102         struct rcu_work rwork;
 103         struct net_device *hw_dev;
 104 };
 105
 106 static const struct rhashtable_params mask_ht_params = {
 107         .key_offset = offsetof(struct fl_flow_mask, key),
 108         .key_len = sizeof(struct fl_flow_key),
 109         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 110         .automatic_shrinking = true,
 111 };
 112
 113 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 114 {
 115         return mask->range.end - mask->range.start;
 116 }
 117
 118 static void fl_mask_update_range(struct fl_flow_mask *mask)
 119 {
 120         const u8 *bytes = (const u8 *) &mask->key;
 121         size_t size = sizeof(mask->key);
 122         size_t i, first = 0, last;
 123
 124         for (i = 0; i < size; i++) {
 125                 if (bytes[i]) {
 126                         first = i;
 127                         break;
 128                 }
 129         }
 130         last = first;
 131         for (i = size - 1; i != first; i--) {
 132                 if (bytes[i]) {
 133                         last = i;
 134                         break;
 135                 }
 136         }
 137         mask->range.start = rounddown(first, sizeof(long));
 138         mask->range.end = roundup(last + 1, sizeof(long));
 139 }
 140
 141 static void *fl_key_get_start(struct fl_flow_key *key,
 142                               const struct fl_flow_mask *mask)
 143 {
 144         return (u8 *) key + mask->range.start;
 145 }
 146
 147 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 148                               struct fl_flow_mask *mask)
 149 {
 150         const long *lkey = fl_key_get_start(key, mask);
 151         const long *lmask = fl_key_get_start(&mask->key, mask);
 152         long *lmkey = fl_key_get_start(mkey, mask);
 153         int i;
 154
 155         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 156                 *lmkey++ = *lkey++ & *lmask++;
 157 }
 158
 159 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 160                                struct fl_flow_mask *mask)
 161 {
 162         const long *lmask = fl_key_get_start(&mask->key, mask);
 163         const long *ltmplt;
 164         int i;
 165
 166         if (!tmplt)
 167                 return true;
 168         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 169         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 170                 if (~*ltmplt++ & *lmask++)
 171                         return false;
 172         }
 173         return true;
 174 }
 175
 176 static void fl_clear_masked_range(struct fl_flow_key *key,
 177                                   struct fl_flow_mask *mask)
 178 {
 179         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 180 }
 181
 182 static struct cls_fl_filter *fl_lookup(struct fl_flow_mask *mask,
 183                                        struct fl_flow_key *mkey)
 184 {
 185         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 186                                       mask->filter_ht_params);
 187 }
 188
 189 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 190                        struct tcf_result *res)
 191 {
 192         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 193         struct cls_fl_filter *f;
 194         struct fl_flow_mask *mask;
 195         struct fl_flow_key skb_key;
 196         struct fl_flow_key skb_mkey;
 197
 198         list_for_each_entry_rcu(mask, &head->masks, list) {
 199                 fl_clear_masked_range(&skb_key, mask);
 200
 201                 skb_key.indev_ifindex = skb->skb_iif;
 202                 /* skb_flow_dissect() does not set n_proto in case an unknown
 203                  * protocol, so do it rather here.
 204                  */
 205                 skb_key.basic.n_proto = skb->protocol;
 206                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 207                 skb_flow_dissect(skb, &mask->dissector, &skb_key, 0);
 208
 209                 fl_set_masked_key(&skb_mkey, &skb_key, mask);
 210
 211                 f = fl_lookup(mask, &skb_mkey);
 212                 if (f && !tc_skip_sw(f->flags)) {
 213                         *res = f->res;
 214                         return tcf_exts_exec(skb, &f->exts, res);
 215                 }
 216         }
 217         return -1;
 218 }
 219
 220 static int fl_init(struct tcf_proto *tp)
 221 {
 222         struct cls_fl_head *head;
 223
 224         head = kzalloc(sizeof(*head), GFP_KERNEL);
 225         if (!head)
 226                 return -ENOBUFS;
 227
 228         INIT_LIST_HEAD_RCU(&head->masks);
 229         rcu_assign_pointer(tp->root, head);
 230         idr_init(&head->handle_idr);
 231
 232         return rhashtable_init(&head->ht, &mask_ht_params);
 233 }
 234
 235 static void fl_mask_free(struct fl_flow_mask *mask)
 236 {
 237         rhashtable_destroy(&mask->ht);
 238         kfree(mask);
 239 }
 240
 241 static void fl_mask_free_work(struct work_struct *work)
 242 {
 243         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 244                                                  struct fl_flow_mask, rwork);
 245
 246         fl_mask_free(mask);
 247 }
 248
 249 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask,
 250                         bool async)
 251 {
 252         if (!list_empty(&mask->filters))
 253                 return false;
 254
 255         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 256         list_del_rcu(&mask->list);
 257         if (async)
 258                 tcf_queue_work(&mask->rwork, fl_mask_free_work);
 259         else
 260                 fl_mask_free(mask);
 261
 262         return true;
 263 }
 264
 265 static void __fl_destroy_filter(struct cls_fl_filter *f)
 266 {
 267         tcf_exts_destroy(&f->exts);
 268         tcf_exts_put_net(&f->exts);
 269         kfree(f);
 270 }
 271
 272 static void fl_destroy_filter_work(struct work_struct *work)
 273 {
 274         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 275                                         struct cls_fl_filter, rwork);
 276
 277         rtnl_lock();
 278         __fl_destroy_filter(f);
 279         rtnl_unlock();
 280 }
 281
 282 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 283                                  struct netlink_ext_ack *extack)
 284 {
 285         struct tc_cls_flower_offload cls_flower = {};
 286         struct tcf_block *block = tp->chain->block;
 287
 288         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 289         cls_flower.command = TC_CLSFLOWER_DESTROY;
 290         cls_flower.cookie = (unsigned long) f;
 291
 292         tc_setup_cb_call(block, &f->exts, TC_SETUP_CLSFLOWER,
 293                          &cls_flower, false);
 294         tcf_block_offload_dec(block, &f->flags);
 295 }
 296
 297 static int fl_hw_replace_filter(struct tcf_proto *tp,
 298                                 struct cls_fl_filter *f,
 299                                 struct netlink_ext_ack *extack)
 300 {
 301         struct tc_cls_flower_offload cls_flower = {};
 302         struct tcf_block *block = tp->chain->block;
 303         bool skip_sw = tc_skip_sw(f->flags);
 304         int err;
 305
 306         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 307         cls_flower.command = TC_CLSFLOWER_REPLACE;
 308         cls_flower.cookie = (unsigned long) f;
 309         cls_flower.dissector = &f->mask->dissector;
 310         cls_flower.mask = &f->mask->key;
 311         cls_flower.key = &f->mkey;
 312         cls_flower.exts = &f->exts;
 313         cls_flower.classid = f->res.classid;
 314
 315         err = tc_setup_cb_call(block, &f->exts, TC_SETUP_CLSFLOWER,
 316                                &cls_flower, skip_sw);
 317         if (err < 0) {
 318                 fl_hw_destroy_filter(tp, f, NULL);
 319                 return err;
 320         } else if (err > 0) {
 321                 f->in_hw_count = err;
 322                 tcf_block_offload_inc(block, &f->flags);
 323         }
 324
 325         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 326                 return -EINVAL;
 327
 328         return 0;
 329 }
 330
 331 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f)
 332 {
 333         struct tc_cls_flower_offload cls_flower = {};
 334         struct tcf_block *block = tp->chain->block;
 335
 336         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 337         cls_flower.command = TC_CLSFLOWER_STATS;
 338         cls_flower.cookie = (unsigned long) f;
 339         cls_flower.exts = &f->exts;
 340         cls_flower.classid = f->res.classid;
 341
 342         tc_setup_cb_call(block, &f->exts, TC_SETUP_CLSFLOWER,
 343                          &cls_flower, false);
 344 }
 345
 346 static bool __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 347                         struct netlink_ext_ack *extack)
 348 {
 349         struct cls_fl_head *head = rtnl_dereference(tp->root);
 350         bool async = tcf_exts_get_net(&f->exts);
 351         bool last;
 352
 353         idr_remove(&head->handle_idr, f->handle);
 354         list_del_rcu(&f->list);
 355         last = fl_mask_put(head, f->mask, async);
 356         if (!tc_skip_hw(f->flags))
 357                 fl_hw_destroy_filter(tp, f, extack);
 358         tcf_unbind_filter(tp, &f->res);
 359         if (async)
 360                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 361         else
 362                 __fl_destroy_filter(f);
 363
 364         return last;
 365 }
 366
 367 static void fl_destroy_sleepable(struct work_struct *work)
 368 {
 369         struct cls_fl_head *head = container_of(to_rcu_work(work),
 370                                                 struct cls_fl_head,
 371                                                 rwork);
 372
 373         rhashtable_destroy(&head->ht);
 374         kfree(head);
 375         module_put(THIS_MODULE);
 376 }
 377
 378 static void fl_destroy(struct tcf_proto *tp, struct netlink_ext_ack *extack)
 379 {
 380         struct cls_fl_head *head = rtnl_dereference(tp->root);
 381         struct fl_flow_mask *mask, *next_mask;
 382         struct cls_fl_filter *f, *next;
 383
 384         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 385                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 386                         if (__fl_delete(tp, f, extack))
 387                                 break;
 388                 }
 389         }
 390         idr_destroy(&head->handle_idr);
 391
 392         __module_get(THIS_MODULE);
 393         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 394 }
 395
 396 static void *fl_get(struct tcf_proto *tp, u32 handle)
 397 {
 398         struct cls_fl_head *head = rtnl_dereference(tp->root);
 399
 400         return idr_find(&head->handle_idr, handle);
 401 }
 402
 403 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 404         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 405         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 406         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 407                                             .len = IFNAMSIZ },
 408         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 409         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 410         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 411         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 412         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 413         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 414         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 415         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 416         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 417         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 418         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 419         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 420         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 421         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 422         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 423         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 424         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 425         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 426         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 427         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 428         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 429         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 430         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 431         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 432         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 433         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 434         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 435         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 436         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 437         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 438         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 439         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 440         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 441         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 442         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 443         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 444         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 445         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 446         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 447         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 448         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 449         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 450         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 451         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 452         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 453         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 454         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 455         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 456         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 457         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 458         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 459         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 460         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 461         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 462         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 463         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 464         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 465         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 466         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 467         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 468         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 469         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 470         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 471         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 472         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 473         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 474         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 475         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 476         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 477         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 478         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 479         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 480         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 481         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 482         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 483         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 484         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 485         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 486         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 487         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 488         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 489 };
 490
 491 static const struct nla_policy
 492 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 493         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 494 };
 495
 496 static const struct nla_policy
 497 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 498         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 499         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 500         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 501                                                        .len = 128 },
 502 };
 503
 504 static void fl_set_key_val(struct nlattr **tb,
 505                            void *val, int val_type,
 506                            void *mask, int mask_type, int len)
 507 {
 508         if (!tb[val_type])
 509                 return;
 510         memcpy(val, nla_data(tb[val_type]), len);
 511         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 512                 memset(mask, 0xff, len);
 513         else
 514                 memcpy(mask, nla_data(tb[mask_type]), len);
 515 }
 516
 517 static int fl_set_key_mpls(struct nlattr **tb,
 518                            struct flow_dissector_key_mpls *key_val,
 519                            struct flow_dissector_key_mpls *key_mask)
 520 {
 521         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 522                 key_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 523                 key_mask->mpls_ttl = MPLS_TTL_MASK;
 524         }
 525         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 526                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 527
 528                 if (bos & ~MPLS_BOS_MASK)
 529                         return -EINVAL;
 530                 key_val->mpls_bos = bos;
 531                 key_mask->mpls_bos = MPLS_BOS_MASK;
 532         }
 533         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 534                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 535
 536                 if (tc & ~MPLS_TC_MASK)
 537                         return -EINVAL;
 538                 key_val->mpls_tc = tc;
 539                 key_mask->mpls_tc = MPLS_TC_MASK;
 540         }
 541         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 542                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 543
 544                 if (label & ~MPLS_LABEL_MASK)
 545                         return -EINVAL;
 546                 key_val->mpls_label = label;
 547                 key_mask->mpls_label = MPLS_LABEL_MASK;
 548         }
 549         return 0;
 550 }
 551
 552 static void fl_set_key_vlan(struct nlattr **tb,
 553                             __be16 ethertype,
 554                             int vlan_id_key, int vlan_prio_key,
 555                             struct flow_dissector_key_vlan *key_val,
 556                             struct flow_dissector_key_vlan *key_mask)
 557 {
 558 #define VLAN_PRIORITY_MASK      0x7
 559
 560         if (tb[vlan_id_key]) {
 561                 key_val->vlan_id =
 562                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
 563                 key_mask->vlan_id = VLAN_VID_MASK;
 564         }
 565         if (tb[vlan_prio_key]) {
 566                 key_val->vlan_priority =
 567                         nla_get_u8(tb[vlan_prio_key]) &
 568                         VLAN_PRIORITY_MASK;
 569                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 570         }
 571         key_val->vlan_tpid = ethertype;
 572         key_mask->vlan_tpid = cpu_to_be16(~0);
 573 }
 574
 575 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
 576                             u32 *dissector_key, u32 *dissector_mask,
 577                             u32 flower_flag_bit, u32 dissector_flag_bit)
 578 {
 579         if (flower_mask & flower_flag_bit) {
 580                 *dissector_mask |= dissector_flag_bit;
 581                 if (flower_key & flower_flag_bit)
 582                         *dissector_key |= dissector_flag_bit;
 583         }
 584 }
 585
 586 static int fl_set_key_flags(struct nlattr **tb,
 587                             u32 *flags_key, u32 *flags_mask)
 588 {
 589         u32 key, mask;
 590
 591         /* mask is mandatory for flags */
 592         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK])
 593                 return -EINVAL;
 594
 595         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
 596         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
 597
 598         *flags_key  = 0;
 599         *flags_mask = 0;
 600
 601         fl_set_key_flag(key, mask, flags_key, flags_mask,
 602                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
 603         fl_set_key_flag(key, mask, flags_key, flags_mask,
 604                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
 605                         FLOW_DIS_FIRST_FRAG);
 606
 607         return 0;
 608 }
 609
 610 static void fl_set_key_ip(struct nlattr **tb, bool encap,
 611                           struct flow_dissector_key_ip *key,
 612                           struct flow_dissector_key_ip *mask)
 613 {
 614         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
 615         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
 616         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
 617         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
 618
 619         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
 620         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
 621 }
 622
 623 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
 624                              int depth, int option_len,
 625                              struct netlink_ext_ack *extack)
 626 {
 627         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
 628         struct nlattr *class = NULL, *type = NULL, *data = NULL;
 629         struct geneve_opt *opt;
 630         int err, data_len = 0;
 631
 632         if (option_len > sizeof(struct geneve_opt))
 633                 data_len = option_len - sizeof(struct geneve_opt);
 634
 635         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
 636         memset(opt, 0xff, option_len);
 637         opt->length = data_len / 4;
 638         opt->r1 = 0;
 639         opt->r2 = 0;
 640         opt->r3 = 0;
 641
 642         /* If no mask has been prodived we assume an exact match. */
 643         if (!depth)
 644                 return sizeof(struct geneve_opt) + data_len;
 645
 646         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
 647                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
 648                 return -EINVAL;
 649         }
 650
 651         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
 652                                nla, geneve_opt_policy, extack);
 653         if (err < 0)
 654                 return err;
 655
 656         /* We are not allowed to omit any of CLASS, TYPE or DATA
 657          * fields from the key.
 658          */
 659         if (!option_len &&
 660             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
 661              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
 662              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
 663                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
 664                 return -EINVAL;
 665         }
 666
 667         /* Omitting any of CLASS, TYPE or DATA fields is allowed
 668          * for the mask.
 669          */
 670         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
 671                 int new_len = key->enc_opts.len;
 672
 673                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
 674                 data_len = nla_len(data);
 675                 if (data_len < 4) {
 676                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
 677                         return -ERANGE;
 678                 }
 679                 if (data_len % 4) {
 680                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
 681                         return -ERANGE;
 682                 }
 683
 684                 new_len += sizeof(struct geneve_opt) + data_len;
 685                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
 686                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
 687                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
 688                         return -ERANGE;
 689                 }
 690                 opt->length = data_len / 4;
 691                 memcpy(opt->opt_data, nla_data(data), data_len);
 692         }
 693
 694         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
 695                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
 696                 opt->opt_class = nla_get_be16(class);
 697         }
 698
 699         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
 700                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
 701                 opt->type = nla_get_u8(type);
 702         }
 703
 704         return sizeof(struct geneve_opt) + data_len;
 705 }
 706
 707 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
 708                           struct fl_flow_key *mask,
 709                           struct netlink_ext_ack *extack)
 710 {
 711         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
 712         int option_len, key_depth, msk_depth = 0;
 713
 714         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
 715
 716         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
 717                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 718                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 719         }
 720
 721         nla_for_each_attr(nla_opt_key, nla_enc_key,
 722                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
 723                 switch (nla_type(nla_opt_key)) {
 724                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
 725                         option_len = 0;
 726                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 727                         option_len = fl_set_geneve_opt(nla_opt_key, key,
 728                                                        key_depth, option_len,
 729                                                        extack);
 730                         if (option_len < 0)
 731                                 return option_len;
 732
 733                         key->enc_opts.len += option_len;
 734                         /* At the same time we need to parse through the mask
 735                          * in order to verify exact and mask attribute lengths.
 736                          */
 737                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 738                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
 739                                                        msk_depth, option_len,
 740                                                        extack);
 741                         if (option_len < 0)
 742                                 return option_len;
 743
 744                         mask->enc_opts.len += option_len;
 745                         if (key->enc_opts.len != mask->enc_opts.len) {
 746                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
 747                                 return -EINVAL;
 748                         }
 749
 750                         if (msk_depth)
 751                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
 752                         break;
 753                 default:
 754                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
 755                         return -EINVAL;
 756                 }
 757         }
 758
 759         return 0;
 760 }
 761
 762 static int fl_set_key(struct net *net, struct nlattr **tb,
 763                       struct fl_flow_key *key, struct fl_flow_key *mask,
 764                       struct netlink_ext_ack *extack)
 765 {
 766         __be16 ethertype;
 767         int ret = 0;
 768 #ifdef CONFIG_NET_CLS_IND
 769         if (tb[TCA_FLOWER_INDEV]) {
 770                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
 771                 if (err < 0)
 772                         return err;
 773                 key->indev_ifindex = err;
 774                 mask->indev_ifindex = 0xffffffff;
 775         }
 776 #endif
 777
 778         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 779                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 780                        sizeof(key->eth.dst));
 781         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 782                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 783                        sizeof(key->eth.src));
 784
 785         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
 786                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
 787
 788                 if (eth_type_vlan(ethertype)) {
 789                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
 790                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
 791                                         &mask->vlan);
 792
 793                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
 794                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
 795                                 if (eth_type_vlan(ethertype)) {
 796                                         fl_set_key_vlan(tb, ethertype,
 797                                                         TCA_FLOWER_KEY_CVLAN_ID,
 798                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
 799                                                         &key->cvlan, &mask->cvlan);
 800                                         fl_set_key_val(tb, &key->basic.n_proto,
 801                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
 802                                                        &mask->basic.n_proto,
 803                                                        TCA_FLOWER_UNSPEC,
 804                                                        sizeof(key->basic.n_proto));
 805                                 } else {
 806                                         key->basic.n_proto = ethertype;
 807                                         mask->basic.n_proto = cpu_to_be16(~0);
 808                                 }
 809                         }
 810                 } else {
 811                         key->basic.n_proto = ethertype;
 812                         mask->basic.n_proto = cpu_to_be16(~0);
 813                 }
 814         }
 815
 816         if (key->basic.n_proto == htons(ETH_P_IP) ||
 817             key->basic.n_proto == htons(ETH_P_IPV6)) {
 818                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 819                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 820                                sizeof(key->basic.ip_proto));
 821                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
 822         }
 823
 824         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
 825                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 826                 mask->control.addr_type = ~0;
 827                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 828                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 829                                sizeof(key->ipv4.src));
 830                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 831                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 832                                sizeof(key->ipv4.dst));
 833         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
 834                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 835                 mask->control.addr_type = ~0;
 836                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 837                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 838                                sizeof(key->ipv6.src));
 839                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 840                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 841                                sizeof(key->ipv6.dst));
 842         }
 843
 844         if (key->basic.ip_proto == IPPROTO_TCP) {
 845                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 846                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 847                                sizeof(key->tp.src));
 848                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 849                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 850                                sizeof(key->tp.dst));
 851                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
 852                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
 853                                sizeof(key->tcp.flags));
 854         } else if (key->basic.ip_proto == IPPROTO_UDP) {
 855                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 856                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 857                                sizeof(key->tp.src));
 858                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 859                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 860                                sizeof(key->tp.dst));
 861         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
 862                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
 863                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
 864                                sizeof(key->tp.src));
 865                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
 866                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
 867                                sizeof(key->tp.dst));
 868         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
 869                    key->basic.ip_proto == IPPROTO_ICMP) {
 870                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
 871                                &mask->icmp.type,
 872                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
 873                                sizeof(key->icmp.type));
 874                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
 875                                &mask->icmp.code,
 876                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
 877                                sizeof(key->icmp.code));
 878         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
 879                    key->basic.ip_proto == IPPROTO_ICMPV6) {
 880                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
 881                                &mask->icmp.type,
 882                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
 883                                sizeof(key->icmp.type));
 884                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
 885                                &mask->icmp.code,
 886                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
 887                                sizeof(key->icmp.code));
 888         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
 889                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
 890                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls);
 891                 if (ret)
 892                         return ret;
 893         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
 894                    key->basic.n_proto == htons(ETH_P_RARP)) {
 895                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
 896                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
 897                                sizeof(key->arp.sip));
 898                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
 899                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
 900                                sizeof(key->arp.tip));
 901                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
 902                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
 903                                sizeof(key->arp.op));
 904                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
 905                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
 906                                sizeof(key->arp.sha));
 907                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
 908                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
 909                                sizeof(key->arp.tha));
 910         }
 911
 912         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
 913             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
 914                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 915                 mask->enc_control.addr_type = ~0;
 916                 fl_set_key_val(tb, &key->enc_ipv4.src,
 917                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
 918                                &mask->enc_ipv4.src,
 919                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 920                                sizeof(key->enc_ipv4.src));
 921                 fl_set_key_val(tb, &key->enc_ipv4.dst,
 922                                TCA_FLOWER_KEY_ENC_IPV4_DST,
 923                                &mask->enc_ipv4.dst,
 924                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 925                                sizeof(key->enc_ipv4.dst));
 926         }
 927
 928         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
 929             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
 930                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 931                 mask->enc_control.addr_type = ~0;
 932                 fl_set_key_val(tb, &key->enc_ipv6.src,
 933                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
 934                                &mask->enc_ipv6.src,
 935                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 936                                sizeof(key->enc_ipv6.src));
 937                 fl_set_key_val(tb, &key->enc_ipv6.dst,
 938                                TCA_FLOWER_KEY_ENC_IPV6_DST,
 939                                &mask->enc_ipv6.dst,
 940                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 941                                sizeof(key->enc_ipv6.dst));
 942         }
 943
 944         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
 945                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
 946                        sizeof(key->enc_key_id.keyid));
 947
 948         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
 949                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
 950                        sizeof(key->enc_tp.src));
 951
 952         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
 953                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
 954                        sizeof(key->enc_tp.dst));
 955
 956         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
 957
 958         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
 959                 ret = fl_set_enc_opt(tb, key, mask, extack);
 960                 if (ret)
 961                         return ret;
 962         }
 963
 964         if (tb[TCA_FLOWER_KEY_FLAGS])
 965                 ret = fl_set_key_flags(tb, &key->control.flags, &mask->control.flags);
 966
 967         return ret;
 968 }
 969
 970 static void fl_mask_copy(struct fl_flow_mask *dst,
 971                          struct fl_flow_mask *src)
 972 {
 973         const void *psrc = fl_key_get_start(&src->key, src);
 974         void *pdst = fl_key_get_start(&dst->key, src);
 975
 976         memcpy(pdst, psrc, fl_mask_range(src));
 977         dst->range = src->range;
 978 }
 979
 980 static const struct rhashtable_params fl_ht_params = {
 981         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
 982         .head_offset = offsetof(struct cls_fl_filter, ht_node),
 983         .automatic_shrinking = true,
 984 };
 985
 986 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
 987 {
 988         mask->filter_ht_params = fl_ht_params;
 989         mask->filter_ht_params.key_len = fl_mask_range(mask);
 990         mask->filter_ht_params.key_offset += mask->range.start;
 991
 992         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
 993 }
 994
 995 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
 996 #define FL_KEY_MEMBER_SIZE(member) (sizeof(((struct fl_flow_key *) 0)->member))
 997
 998 #define FL_KEY_IS_MASKED(mask, member)                                          \
 999         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1000                    0, FL_KEY_MEMBER_SIZE(member))                               \
1001
1002 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1003         do {                                                                    \
1004                 keys[cnt].key_id = id;                                          \
1005                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1006                 cnt++;                                                          \
1007         } while(0);
1008
1009 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1010         do {                                                                    \
1011                 if (FL_KEY_IS_MASKED(mask, member))                             \
1012                         FL_KEY_SET(keys, cnt, id, member);                      \
1013         } while(0);
1014
1015 static void fl_init_dissector(struct flow_dissector *dissector,
1016                               struct fl_flow_key *mask)
1017 {
1018         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1019         size_t cnt = 0;
1020
1021         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1022         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1023         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1024                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1025         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1026                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1027         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1028                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1029         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1030                              FLOW_DISSECTOR_KEY_PORTS, tp);
1031         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1032                              FLOW_DISSECTOR_KEY_IP, ip);
1033         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1034                              FLOW_DISSECTOR_KEY_TCP, tcp);
1035         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1036                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1037         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1038                              FLOW_DISSECTOR_KEY_ARP, arp);
1039         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1040                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1041         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1042                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1043         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1044                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1045         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1046                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1047         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1048                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1049         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1050                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1051         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1052             FL_KEY_IS_MASKED(mask, enc_ipv6))
1053                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1054                            enc_control);
1055         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1056                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1057         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1058                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1059         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1060                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1061
1062         skb_flow_dissector_init(dissector, keys, cnt);
1063 }
1064
1065 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1066                                                struct fl_flow_mask *mask)
1067 {
1068         struct fl_flow_mask *newmask;
1069         int err;
1070
1071         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1072         if (!newmask)
1073                 return ERR_PTR(-ENOMEM);
1074
1075         fl_mask_copy(newmask, mask);
1076
1077         err = fl_init_mask_hashtable(newmask);
1078         if (err)
1079                 goto errout_free;
1080
1081         fl_init_dissector(&newmask->dissector, &newmask->key);
1082
1083         INIT_LIST_HEAD_RCU(&newmask->filters);
1084
1085         err = rhashtable_insert_fast(&head->ht, &newmask->ht_node,
1086                                      mask_ht_params);
1087         if (err)
1088                 goto errout_destroy;
1089
1090         list_add_tail_rcu(&newmask->list, &head->masks);
1091
1092         return newmask;
1093
1094 errout_destroy:
1095         rhashtable_destroy(&newmask->ht);
1096 errout_free:
1097         kfree(newmask);
1098
1099         return ERR_PTR(err);
1100 }
1101
1102 static int fl_check_assign_mask(struct cls_fl_head *head,
1103                                 struct cls_fl_filter *fnew,
1104                                 struct cls_fl_filter *fold,
1105                                 struct fl_flow_mask *mask)
1106 {
1107         struct fl_flow_mask *newmask;
1108
1109         fnew->mask = rhashtable_lookup_fast(&head->ht, mask, mask_ht_params);
1110         if (!fnew->mask) {
1111                 if (fold)
1112                         return -EINVAL;
1113
1114                 newmask = fl_create_new_mask(head, mask);
1115                 if (IS_ERR(newmask))
1116                         return PTR_ERR(newmask);
1117
1118                 fnew->mask = newmask;
1119         } else if (fold && fold->mask != fnew->mask) {
1120                 return -EINVAL;
1121         }
1122
1123         return 0;
1124 }
1125
1126 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1127                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1128                         unsigned long base, struct nlattr **tb,
1129                         struct nlattr *est, bool ovr,
1130                         struct fl_flow_tmplt *tmplt,
1131                         struct netlink_ext_ack *extack)
1132 {
1133         int err;
1134
1135         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, extack);
1136         if (err < 0)
1137                 return err;
1138
1139         if (tb[TCA_FLOWER_CLASSID]) {
1140                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1141                 tcf_bind_filter(tp, &f->res, base);
1142         }
1143
1144         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1145         if (err)
1146                 return err;
1147
1148         fl_mask_update_range(mask);
1149         fl_set_masked_key(&f->mkey, &f->key, mask);
1150
1151         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1152                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1153                 return -EINVAL;
1154         }
1155
1156         return 0;
1157 }
1158
1159 static int fl_change(struct net *net, struct sk_buff *in_skb,
1160                      struct tcf_proto *tp, unsigned long base,
1161                      u32 handle, struct nlattr **tca,
1162                      void **arg, bool ovr, struct netlink_ext_ack *extack)
1163 {
1164         struct cls_fl_head *head = rtnl_dereference(tp->root);
1165         struct cls_fl_filter *fold = *arg;
1166         struct cls_fl_filter *fnew;
1167         struct nlattr **tb;
1168         struct fl_flow_mask mask = {};
1169         int err;
1170
1171         if (!tca[TCA_OPTIONS])
1172                 return -EINVAL;
1173
1174         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1175         if (!tb)
1176                 return -ENOBUFS;
1177
1178         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS],
1179                                fl_policy, NULL);
1180         if (err < 0)
1181                 goto errout_tb;
1182
1183         if (fold && handle && fold->handle != handle) {
1184                 err = -EINVAL;
1185                 goto errout_tb;
1186         }
1187
1188         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
1189         if (!fnew) {
1190                 err = -ENOBUFS;
1191                 goto errout_tb;
1192         }
1193
1194         err = tcf_exts_init(&fnew->exts, TCA_FLOWER_ACT, 0);
1195         if (err < 0)
1196                 goto errout;
1197
1198         if (!handle) {
1199                 handle = 1;
1200                 err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1201                                     INT_MAX, GFP_KERNEL);
1202         } else if (!fold) {
1203                 /* user specifies a handle and it doesn't exist */
1204                 err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1205                                     handle, GFP_KERNEL);
1206         }
1207         if (err)
1208                 goto errout;
1209         fnew->handle = handle;
1210
1211         if (tb[TCA_FLOWER_FLAGS]) {
1212                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
1213
1214                 if (!tc_flags_valid(fnew->flags)) {
1215                         err = -EINVAL;
1216                         goto errout_idr;
1217                 }
1218         }
1219
1220         err = fl_set_parms(net, tp, fnew, &mask, base, tb, tca[TCA_RATE], ovr,
1221                            tp->chain->tmplt_priv, extack);
1222         if (err)
1223                 goto errout_idr;
1224
1225         err = fl_check_assign_mask(head, fnew, fold, &mask);
1226         if (err)
1227                 goto errout_idr;
1228
1229         if (!tc_skip_sw(fnew->flags)) {
1230                 if (!fold && fl_lookup(fnew->mask, &fnew->mkey)) {
1231                         err = -EEXIST;
1232                         goto errout_mask;
1233                 }
1234
1235                 err = rhashtable_insert_fast(&fnew->mask->ht, &fnew->ht_node,
1236                                              fnew->mask->filter_ht_params);
1237                 if (err)
1238                         goto errout_mask;
1239         }
1240
1241         if (!tc_skip_hw(fnew->flags)) {
1242                 err = fl_hw_replace_filter(tp, fnew, extack);
1243                 if (err)
1244                         goto errout_mask;
1245         }
1246
1247         if (!tc_in_hw(fnew->flags))
1248                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1249
1250         if (fold) {
1251                 if (!tc_skip_sw(fold->flags))
1252                         rhashtable_remove_fast(&fold->mask->ht,
1253                                                &fold->ht_node,
1254                                                fold->mask->filter_ht_params);
1255                 if (!tc_skip_hw(fold->flags))
1256                         fl_hw_destroy_filter(tp, fold, NULL);
1257         }
1258
1259         *arg = fnew;
1260
1261         if (fold) {
1262                 idr_replace(&head->handle_idr, fnew, fnew->handle);
1263                 list_replace_rcu(&fold->list, &fnew->list);
1264                 tcf_unbind_filter(tp, &fold->res);
1265                 tcf_exts_get_net(&fold->exts);
1266                 tcf_queue_work(&fold->rwork, fl_destroy_filter_work);
1267         } else {
1268                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
1269         }
1270
1271         kfree(tb);
1272         return 0;
1273
1274 errout_mask:
1275         fl_mask_put(head, fnew->mask, false);
1276
1277 errout_idr:
1278         if (!fold)
1279                 idr_remove(&head->handle_idr, fnew->handle);
1280 errout:
1281         tcf_exts_destroy(&fnew->exts);
1282         kfree(fnew);
1283 errout_tb:
1284         kfree(tb);
1285         return err;
1286 }
1287
1288 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
1289                      struct netlink_ext_ack *extack)
1290 {
1291         struct cls_fl_head *head = rtnl_dereference(tp->root);
1292         struct cls_fl_filter *f = arg;
1293
1294         if (!tc_skip_sw(f->flags))
1295                 rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
1296                                        f->mask->filter_ht_params);
1297         __fl_delete(tp, f, extack);
1298         *last = list_empty(&head->masks);
1299         return 0;
1300 }
1301
1302 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg)
1303 {
1304         struct cls_fl_head *head = rtnl_dereference(tp->root);
1305         struct cls_fl_filter *f;
1306
1307         arg->count = arg->skip;
1308
1309         while ((f = idr_get_next_ul(&head->handle_idr,
1310                                     &arg->cookie)) != NULL) {
1311                 if (arg->fn(tp, f, arg) < 0) {
1312                         arg->stop = 1;
1313                         break;
1314                 }
1315                 arg->cookie = f->handle + 1;
1316                 arg->count++;
1317         }
1318 }
1319
1320 static int fl_reoffload(struct tcf_proto *tp, bool add, tc_setup_cb_t *cb,
1321                         void *cb_priv, struct netlink_ext_ack *extack)
1322 {
1323         struct cls_fl_head *head = rtnl_dereference(tp->root);
1324         struct tc_cls_flower_offload cls_flower = {};
1325         struct tcf_block *block = tp->chain->block;
1326         struct fl_flow_mask *mask;
1327         struct cls_fl_filter *f;
1328         int err;
1329
1330         list_for_each_entry(mask, &head->masks, list) {
1331                 list_for_each_entry(f, &mask->filters, list) {
1332                         if (tc_skip_hw(f->flags))
1333                                 continue;
1334
1335                         tc_cls_common_offload_init(&cls_flower.common, tp,
1336                                                    f->flags, extack);
1337                         cls_flower.command = add ?
1338                                 TC_CLSFLOWER_REPLACE : TC_CLSFLOWER_DESTROY;
1339                         cls_flower.cookie = (unsigned long)f;
1340                         cls_flower.dissector = &mask->dissector;
1341                         cls_flower.mask = &mask->key;
1342                         cls_flower.key = &f->mkey;
1343                         cls_flower.exts = &f->exts;
1344                         cls_flower.classid = f->res.classid;
1345
1346                         err = cb(TC_SETUP_CLSFLOWER, &cls_flower, cb_priv);
1347                         if (err) {
1348                                 if (add && tc_skip_sw(f->flags))
1349                                         return err;
1350                                 continue;
1351                         }
1352
1353                         tc_cls_offload_cnt_update(block, &f->in_hw_count,
1354                                                   &f->flags, add);
1355                 }
1356         }
1357
1358         return 0;
1359 }
1360
1361 static void fl_hw_create_tmplt(struct tcf_chain *chain,
1362                                struct fl_flow_tmplt *tmplt)
1363 {
1364         struct tc_cls_flower_offload cls_flower = {};
1365         struct tcf_block *block = chain->block;
1366         struct tcf_exts dummy_exts = { 0, };
1367
1368         cls_flower.common.chain_index = chain->index;
1369         cls_flower.command = TC_CLSFLOWER_TMPLT_CREATE;
1370         cls_flower.cookie = (unsigned long) tmplt;
1371         cls_flower.dissector = &tmplt->dissector;
1372         cls_flower.mask = &tmplt->mask;
1373         cls_flower.key = &tmplt->dummy_key;
1374         cls_flower.exts = &dummy_exts;
1375
1376         /* We don't care if driver (any of them) fails to handle this
1377          * call. It serves just as a hint for it.
1378          */
1379         tc_setup_cb_call(block, NULL, TC_SETUP_CLSFLOWER,
1380                          &cls_flower, false);
1381 }
1382
1383 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
1384                                 struct fl_flow_tmplt *tmplt)
1385 {
1386         struct tc_cls_flower_offload cls_flower = {};
1387         struct tcf_block *block = chain->block;
1388
1389         cls_flower.common.chain_index = chain->index;
1390         cls_flower.command = TC_CLSFLOWER_TMPLT_DESTROY;
1391         cls_flower.cookie = (unsigned long) tmplt;
1392
1393         tc_setup_cb_call(block, NULL, TC_SETUP_CLSFLOWER,
1394                          &cls_flower, false);
1395 }
1396
1397 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
1398                              struct nlattr **tca,
1399                              struct netlink_ext_ack *extack)
1400 {
1401         struct fl_flow_tmplt *tmplt;
1402         struct nlattr **tb;
1403         int err;
1404
1405         if (!tca[TCA_OPTIONS])
1406                 return ERR_PTR(-EINVAL);
1407
1408         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1409         if (!tb)
1410                 return ERR_PTR(-ENOBUFS);
1411         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS],
1412                                fl_policy, NULL);
1413         if (err)
1414                 goto errout_tb;
1415
1416         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
1417         if (!tmplt) {
1418                 err = -ENOMEM;
1419                 goto errout_tb;
1420         }
1421         tmplt->chain = chain;
1422         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
1423         if (err)
1424                 goto errout_tmplt;
1425         kfree(tb);
1426
1427         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
1428
1429         fl_hw_create_tmplt(chain, tmplt);
1430
1431         return tmplt;
1432
1433 errout_tmplt:
1434         kfree(tmplt);
1435 errout_tb:
1436         kfree(tb);
1437         return ERR_PTR(err);
1438 }
1439
1440 static void fl_tmplt_destroy(void *tmplt_priv)
1441 {
1442         struct fl_flow_tmplt *tmplt = tmplt_priv;
1443
1444         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
1445         kfree(tmplt);
1446 }
1447
1448 static int fl_dump_key_val(struct sk_buff *skb,
1449                            void *val, int val_type,
1450                            void *mask, int mask_type, int len)
1451 {
1452         int err;
1453
1454         if (!memchr_inv(mask, 0, len))
1455                 return 0;
1456         err = nla_put(skb, val_type, len, val);
1457         if (err)
1458                 return err;
1459         if (mask_type != TCA_FLOWER_UNSPEC) {
1460                 err = nla_put(skb, mask_type, len, mask);
1461                 if (err)
1462                         return err;
1463         }
1464         return 0;
1465 }
1466
1467 static int fl_dump_key_mpls(struct sk_buff *skb,
1468                             struct flow_dissector_key_mpls *mpls_key,
1469                             struct flow_dissector_key_mpls *mpls_mask)
1470 {
1471         int err;
1472
1473         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
1474                 return 0;
1475         if (mpls_mask->mpls_ttl) {
1476                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
1477                                  mpls_key->mpls_ttl);
1478                 if (err)
1479                         return err;
1480         }
1481         if (mpls_mask->mpls_tc) {
1482                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
1483                                  mpls_key->mpls_tc);
1484                 if (err)
1485                         return err;
1486         }
1487         if (mpls_mask->mpls_label) {
1488                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
1489                                   mpls_key->mpls_label);
1490                 if (err)
1491                         return err;
1492         }
1493         if (mpls_mask->mpls_bos) {
1494                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
1495                                  mpls_key->mpls_bos);
1496                 if (err)
1497                         return err;
1498         }
1499         return 0;
1500 }
1501
1502 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
1503                           struct flow_dissector_key_ip *key,
1504                           struct flow_dissector_key_ip *mask)
1505 {
1506         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1507         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1508         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1509         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1510
1511         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
1512             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
1513                 return -1;
1514
1515         return 0;
1516 }
1517
1518 static int fl_dump_key_vlan(struct sk_buff *skb,
1519                             int vlan_id_key, int vlan_prio_key,
1520                             struct flow_dissector_key_vlan *vlan_key,
1521                             struct flow_dissector_key_vlan *vlan_mask)
1522 {
1523         int err;
1524
1525         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
1526                 return 0;
1527         if (vlan_mask->vlan_id) {
1528                 err = nla_put_u16(skb, vlan_id_key,
1529                                   vlan_key->vlan_id);
1530                 if (err)
1531                         return err;
1532         }
1533         if (vlan_mask->vlan_priority) {
1534                 err = nla_put_u8(skb, vlan_prio_key,
1535                                  vlan_key->vlan_priority);
1536                 if (err)
1537                         return err;
1538         }
1539         return 0;
1540 }
1541
1542 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
1543                             u32 *flower_key, u32 *flower_mask,
1544                             u32 flower_flag_bit, u32 dissector_flag_bit)
1545 {
1546         if (dissector_mask & dissector_flag_bit) {
1547                 *flower_mask |= flower_flag_bit;
1548                 if (dissector_key & dissector_flag_bit)
1549                         *flower_key |= flower_flag_bit;
1550         }
1551 }
1552
1553 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
1554 {
1555         u32 key, mask;
1556         __be32 _key, _mask;
1557         int err;
1558
1559         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
1560                 return 0;
1561
1562         key = 0;
1563         mask = 0;
1564
1565         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
1566                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1567         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
1568                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1569                         FLOW_DIS_FIRST_FRAG);
1570
1571         _key = cpu_to_be32(key);
1572         _mask = cpu_to_be32(mask);
1573
1574         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
1575         if (err)
1576                 return err;
1577
1578         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
1579 }
1580
1581 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
1582                                   struct flow_dissector_key_enc_opts *enc_opts)
1583 {
1584         struct geneve_opt *opt;
1585         struct nlattr *nest;
1586         int opt_off = 0;
1587
1588         nest = nla_nest_start(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
1589         if (!nest)
1590                 goto nla_put_failure;
1591
1592         while (enc_opts->len > opt_off) {
1593                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
1594
1595                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
1596                                  opt->opt_class))
1597                         goto nla_put_failure;
1598                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
1599                                opt->type))
1600                         goto nla_put_failure;
1601                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
1602                             opt->length * 4, opt->opt_data))
1603                         goto nla_put_failure;
1604
1605                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
1606         }
1607         nla_nest_end(skb, nest);
1608         return 0;
1609
1610 nla_put_failure:
1611         nla_nest_cancel(skb, nest);
1612         return -EMSGSIZE;
1613 }
1614
1615 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
1616                                struct flow_dissector_key_enc_opts *enc_opts)
1617 {
1618         struct nlattr *nest;
1619         int err;
1620
1621         if (!enc_opts->len)
1622                 return 0;
1623
1624         nest = nla_nest_start(skb, enc_opt_type);
1625         if (!nest)
1626                 goto nla_put_failure;
1627
1628         switch (enc_opts->dst_opt_type) {
1629         case TUNNEL_GENEVE_OPT:
1630                 err = fl_dump_key_geneve_opt(skb, enc_opts);
1631                 if (err)
1632                         goto nla_put_failure;
1633                 break;
1634         default:
1635                 goto nla_put_failure;
1636         }
1637         nla_nest_end(skb, nest);
1638         return 0;
1639
1640 nla_put_failure:
1641         nla_nest_cancel(skb, nest);
1642         return -EMSGSIZE;
1643 }
1644
1645 static int fl_dump_key_enc_opt(struct sk_buff *skb,
1646                                struct flow_dissector_key_enc_opts *key_opts,
1647                                struct flow_dissector_key_enc_opts *msk_opts)
1648 {
1649         int err;
1650
1651         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
1652         if (err)
1653                 return err;
1654
1655         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
1656 }
1657
1658 static int fl_dump_key(struct sk_buff *skb, struct net *net,
1659                        struct fl_flow_key *key, struct fl_flow_key *mask)
1660 {
1661         if (mask->indev_ifindex) {
1662                 struct net_device *dev;
1663
1664                 dev = __dev_get_by_index(net, key->indev_ifindex);
1665                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
1666                         goto nla_put_failure;
1667         }
1668
1669         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1670                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1671                             sizeof(key->eth.dst)) ||
1672             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1673                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1674                             sizeof(key->eth.src)) ||
1675             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
1676                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
1677                             sizeof(key->basic.n_proto)))
1678                 goto nla_put_failure;
1679
1680         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
1681                 goto nla_put_failure;
1682
1683         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
1684                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
1685                 goto nla_put_failure;
1686
1687         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
1688                              TCA_FLOWER_KEY_CVLAN_PRIO,
1689                              &key->cvlan, &mask->cvlan) ||
1690             (mask->cvlan.vlan_tpid &&
1691              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1692                           key->cvlan.vlan_tpid)))
1693                 goto nla_put_failure;
1694
1695         if (mask->basic.n_proto) {
1696                 if (mask->cvlan.vlan_tpid) {
1697                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1698                                          key->basic.n_proto))
1699                                 goto nla_put_failure;
1700                 } else if (mask->vlan.vlan_tpid) {
1701                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1702                                          key->basic.n_proto))
1703                                 goto nla_put_failure;
1704                 }
1705         }
1706
1707         if ((key->basic.n_proto == htons(ETH_P_IP) ||
1708              key->basic.n_proto == htons(ETH_P_IPV6)) &&
1709             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1710                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1711                             sizeof(key->basic.ip_proto)) ||
1712             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
1713                 goto nla_put_failure;
1714
1715         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
1716             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1717                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1718                              sizeof(key->ipv4.src)) ||
1719              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1720                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1721                              sizeof(key->ipv4.dst))))
1722                 goto nla_put_failure;
1723         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
1724                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1725                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1726                                   sizeof(key->ipv6.src)) ||
1727                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1728                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1729                                   sizeof(key->ipv6.dst))))
1730                 goto nla_put_failure;
1731
1732         if (key->basic.ip_proto == IPPROTO_TCP &&
1733             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1734                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1735                              sizeof(key->tp.src)) ||
1736              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1737                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1738                              sizeof(key->tp.dst)) ||
1739              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1740                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1741                              sizeof(key->tcp.flags))))
1742                 goto nla_put_failure;
1743         else if (key->basic.ip_proto == IPPROTO_UDP &&
1744                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1745                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1746                                   sizeof(key->tp.src)) ||
1747                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1748                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1749                                   sizeof(key->tp.dst))))
1750                 goto nla_put_failure;
1751         else if (key->basic.ip_proto == IPPROTO_SCTP &&
1752                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1753                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1754                                   sizeof(key->tp.src)) ||
1755                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1756                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1757                                   sizeof(key->tp.dst))))
1758                 goto nla_put_failure;
1759         else if (key->basic.n_proto == htons(ETH_P_IP) &&
1760                  key->basic.ip_proto == IPPROTO_ICMP &&
1761                  (fl_dump_key_val(skb, &key->icmp.type,
1762                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
1763                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1764                                   sizeof(key->icmp.type)) ||
1765                   fl_dump_key_val(skb, &key->icmp.code,
1766                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
1767                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1768                                   sizeof(key->icmp.code))))
1769                 goto nla_put_failure;
1770         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1771                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
1772                  (fl_dump_key_val(skb, &key->icmp.type,
1773                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
1774                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1775                                   sizeof(key->icmp.type)) ||
1776                   fl_dump_key_val(skb, &key->icmp.code,
1777                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
1778                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1779                                   sizeof(key->icmp.code))))
1780                 goto nla_put_failure;
1781         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
1782                   key->basic.n_proto == htons(ETH_P_RARP)) &&
1783                  (fl_dump_key_val(skb, &key->arp.sip,
1784                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
1785                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
1786                                   sizeof(key->arp.sip)) ||
1787                   fl_dump_key_val(skb, &key->arp.tip,
1788                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
1789                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
1790                                   sizeof(key->arp.tip)) ||
1791                   fl_dump_key_val(skb, &key->arp.op,
1792                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
1793                                   TCA_FLOWER_KEY_ARP_OP_MASK,
1794                                   sizeof(key->arp.op)) ||
1795                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1796                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1797                                   sizeof(key->arp.sha)) ||
1798                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1799                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1800                                   sizeof(key->arp.tha))))
1801                 goto nla_put_failure;
1802
1803         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
1804             (fl_dump_key_val(skb, &key->enc_ipv4.src,
1805                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
1806                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1807                             sizeof(key->enc_ipv4.src)) ||
1808              fl_dump_key_val(skb, &key->enc_ipv4.dst,
1809                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
1810                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1811                              sizeof(key->enc_ipv4.dst))))
1812                 goto nla_put_failure;
1813         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
1814                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
1815                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
1816                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1817                             sizeof(key->enc_ipv6.src)) ||
1818                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
1819                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
1820                                  &mask->enc_ipv6.dst,
1821                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1822                             sizeof(key->enc_ipv6.dst))))
1823                 goto nla_put_failure;
1824
1825         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
1826                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
1827                             sizeof(key->enc_key_id)) ||
1828             fl_dump_key_val(skb, &key->enc_tp.src,
1829                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1830                             &mask->enc_tp.src,
1831                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1832                             sizeof(key->enc_tp.src)) ||
1833             fl_dump_key_val(skb, &key->enc_tp.dst,
1834                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1835                             &mask->enc_tp.dst,
1836                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1837                             sizeof(key->enc_tp.dst)) ||
1838             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
1839             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
1840                 goto nla_put_failure;
1841
1842         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
1843                 goto nla_put_failure;
1844
1845         return 0;
1846
1847 nla_put_failure:
1848         return -EMSGSIZE;
1849 }
1850
1851 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
1852                    struct sk_buff *skb, struct tcmsg *t)
1853 {
1854         struct cls_fl_filter *f = fh;
1855         struct nlattr *nest;
1856         struct fl_flow_key *key, *mask;
1857
1858         if (!f)
1859                 return skb->len;
1860
1861         t->tcm_handle = f->handle;
1862
1863         nest = nla_nest_start(skb, TCA_OPTIONS);
1864         if (!nest)
1865                 goto nla_put_failure;
1866
1867         if (f->res.classid &&
1868             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
1869                 goto nla_put_failure;
1870
1871         key = &f->key;
1872         mask = &f->mask->key;
1873
1874         if (fl_dump_key(skb, net, key, mask))
1875                 goto nla_put_failure;
1876
1877         if (!tc_skip_hw(f->flags))
1878                 fl_hw_update_stats(tp, f);
1879
1880         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
1881                 goto nla_put_failure;
1882
1883         if (tcf_exts_dump(skb, &f->exts))
1884                 goto nla_put_failure;
1885
1886         nla_nest_end(skb, nest);
1887
1888         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
1889                 goto nla_put_failure;
1890
1891         return skb->len;
1892
1893 nla_put_failure:
1894         nla_nest_cancel(skb, nest);
1895         return -1;
1896 }
1897
1898 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
1899 {
1900         struct fl_flow_tmplt *tmplt = tmplt_priv;
1901         struct fl_flow_key *key, *mask;
1902         struct nlattr *nest;
1903
1904         nest = nla_nest_start(skb, TCA_OPTIONS);
1905         if (!nest)
1906                 goto nla_put_failure;
1907
1908         key = &tmplt->dummy_key;
1909         mask = &tmplt->mask;
1910
1911         if (fl_dump_key(skb, net, key, mask))
1912                 goto nla_put_failure;
1913
1914         nla_nest_end(skb, nest);
1915
1916         return skb->len;
1917
1918 nla_put_failure:
1919         nla_nest_cancel(skb, nest);
1920         return -EMSGSIZE;
1921 }
1922
1923 static void fl_bind_class(void *fh, u32 classid, unsigned long cl)
1924 {
1925         struct cls_fl_filter *f = fh;
1926
1927         if (f && f->res.classid == classid)
1928                 f->res.class = cl;
1929 }
1930
1931 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
1932         .kind           = "flower",
1933         .classify       = fl_classify,
1934         .init           = fl_init,
1935         .destroy        = fl_destroy,
1936         .get            = fl_get,
1937         .change         = fl_change,
1938         .delete         = fl_delete,
1939         .walk           = fl_walk,
1940         .reoffload      = fl_reoffload,
1941         .dump           = fl_dump,
1942         .bind_class     = fl_bind_class,
1943         .tmplt_create   = fl_tmplt_create,
1944         .tmplt_destroy  = fl_tmplt_destroy,
1945         .tmplt_dump     = fl_tmplt_dump,
1946         .owner          = THIS_MODULE,
1947 };
1948
1949 static int __init cls_fl_init(void)
1950 {
1951         return register_tcf_proto_ops(&cls_fl_ops);
1952 }
1953
1954 static void __exit cls_fl_exit(void)
1955 {
1956         unregister_tcf_proto_ops(&cls_fl_ops);
1957 }
1958
1959 module_init(cls_fl_init);
1960 module_exit(cls_fl_exit);
1961
1962 MODULE_AUTHOR("Jiri Pirko <[email protected]>");
1963 MODULE_DESCRIPTION("Flower classifier");
1964 MODULE_LICENSE("GPL v2");