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
  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         u32 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 err, option_len, key_depth, msk_depth = 0;
 713
 714         err = nla_validate_nested(tb[TCA_FLOWER_KEY_ENC_OPTS],
 715                                   TCA_FLOWER_KEY_ENC_OPTS_MAX,
 716                                   enc_opts_policy, extack);
 717         if (err)
 718                 return err;
 719
 720         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
 721
 722         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
 723                 err = nla_validate_nested(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
 724                                           TCA_FLOWER_KEY_ENC_OPTS_MAX,
 725                                           enc_opts_policy, extack);
 726                 if (err)
 727                         return err;
 728
 729                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 730                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 731         }
 732
 733         nla_for_each_attr(nla_opt_key, nla_enc_key,
 734                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
 735                 switch (nla_type(nla_opt_key)) {
 736                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
 737                         option_len = 0;
 738                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 739                         option_len = fl_set_geneve_opt(nla_opt_key, key,
 740                                                        key_depth, option_len,
 741                                                        extack);
 742                         if (option_len < 0)
 743                                 return option_len;
 744
 745                         key->enc_opts.len += option_len;
 746                         /* At the same time we need to parse through the mask
 747                          * in order to verify exact and mask attribute lengths.
 748                          */
 749                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 750                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
 751                                                        msk_depth, option_len,
 752                                                        extack);
 753                         if (option_len < 0)
 754                                 return option_len;
 755
 756                         mask->enc_opts.len += option_len;
 757                         if (key->enc_opts.len != mask->enc_opts.len) {
 758                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
 759                                 return -EINVAL;
 760                         }
 761
 762                         if (msk_depth)
 763                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
 764                         break;
 765                 default:
 766                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
 767                         return -EINVAL;
 768                 }
 769         }
 770
 771         return 0;
 772 }
 773
 774 static int fl_set_key(struct net *net, struct nlattr **tb,
 775                       struct fl_flow_key *key, struct fl_flow_key *mask,
 776                       struct netlink_ext_ack *extack)
 777 {
 778         __be16 ethertype;
 779         int ret = 0;
 780 #ifdef CONFIG_NET_CLS_IND
 781         if (tb[TCA_FLOWER_INDEV]) {
 782                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
 783                 if (err < 0)
 784                         return err;
 785                 key->indev_ifindex = err;
 786                 mask->indev_ifindex = 0xffffffff;
 787         }
 788 #endif
 789
 790         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 791                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 792                        sizeof(key->eth.dst));
 793         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 794                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 795                        sizeof(key->eth.src));
 796
 797         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
 798                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
 799
 800                 if (eth_type_vlan(ethertype)) {
 801                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
 802                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
 803                                         &mask->vlan);
 804
 805                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
 806                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
 807                                 if (eth_type_vlan(ethertype)) {
 808                                         fl_set_key_vlan(tb, ethertype,
 809                                                         TCA_FLOWER_KEY_CVLAN_ID,
 810                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
 811                                                         &key->cvlan, &mask->cvlan);
 812                                         fl_set_key_val(tb, &key->basic.n_proto,
 813                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
 814                                                        &mask->basic.n_proto,
 815                                                        TCA_FLOWER_UNSPEC,
 816                                                        sizeof(key->basic.n_proto));
 817                                 } else {
 818                                         key->basic.n_proto = ethertype;
 819                                         mask->basic.n_proto = cpu_to_be16(~0);
 820                                 }
 821                         }
 822                 } else {
 823                         key->basic.n_proto = ethertype;
 824                         mask->basic.n_proto = cpu_to_be16(~0);
 825                 }
 826         }
 827
 828         if (key->basic.n_proto == htons(ETH_P_IP) ||
 829             key->basic.n_proto == htons(ETH_P_IPV6)) {
 830                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 831                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 832                                sizeof(key->basic.ip_proto));
 833                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
 834         }
 835
 836         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
 837                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 838                 mask->control.addr_type = ~0;
 839                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 840                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 841                                sizeof(key->ipv4.src));
 842                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 843                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 844                                sizeof(key->ipv4.dst));
 845         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
 846                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 847                 mask->control.addr_type = ~0;
 848                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 849                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 850                                sizeof(key->ipv6.src));
 851                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 852                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 853                                sizeof(key->ipv6.dst));
 854         }
 855
 856         if (key->basic.ip_proto == IPPROTO_TCP) {
 857                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 858                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 859                                sizeof(key->tp.src));
 860                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 861                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 862                                sizeof(key->tp.dst));
 863                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
 864                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
 865                                sizeof(key->tcp.flags));
 866         } else if (key->basic.ip_proto == IPPROTO_UDP) {
 867                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 868                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 869                                sizeof(key->tp.src));
 870                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 871                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 872                                sizeof(key->tp.dst));
 873         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
 874                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
 875                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
 876                                sizeof(key->tp.src));
 877                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
 878                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
 879                                sizeof(key->tp.dst));
 880         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
 881                    key->basic.ip_proto == IPPROTO_ICMP) {
 882                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
 883                                &mask->icmp.type,
 884                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
 885                                sizeof(key->icmp.type));
 886                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
 887                                &mask->icmp.code,
 888                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
 889                                sizeof(key->icmp.code));
 890         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
 891                    key->basic.ip_proto == IPPROTO_ICMPV6) {
 892                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
 893                                &mask->icmp.type,
 894                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
 895                                sizeof(key->icmp.type));
 896                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
 897                                &mask->icmp.code,
 898                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
 899                                sizeof(key->icmp.code));
 900         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
 901                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
 902                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls);
 903                 if (ret)
 904                         return ret;
 905         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
 906                    key->basic.n_proto == htons(ETH_P_RARP)) {
 907                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
 908                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
 909                                sizeof(key->arp.sip));
 910                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
 911                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
 912                                sizeof(key->arp.tip));
 913                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
 914                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
 915                                sizeof(key->arp.op));
 916                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
 917                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
 918                                sizeof(key->arp.sha));
 919                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
 920                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
 921                                sizeof(key->arp.tha));
 922         }
 923
 924         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
 925             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
 926                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 927                 mask->enc_control.addr_type = ~0;
 928                 fl_set_key_val(tb, &key->enc_ipv4.src,
 929                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
 930                                &mask->enc_ipv4.src,
 931                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 932                                sizeof(key->enc_ipv4.src));
 933                 fl_set_key_val(tb, &key->enc_ipv4.dst,
 934                                TCA_FLOWER_KEY_ENC_IPV4_DST,
 935                                &mask->enc_ipv4.dst,
 936                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 937                                sizeof(key->enc_ipv4.dst));
 938         }
 939
 940         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
 941             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
 942                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 943                 mask->enc_control.addr_type = ~0;
 944                 fl_set_key_val(tb, &key->enc_ipv6.src,
 945                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
 946                                &mask->enc_ipv6.src,
 947                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 948                                sizeof(key->enc_ipv6.src));
 949                 fl_set_key_val(tb, &key->enc_ipv6.dst,
 950                                TCA_FLOWER_KEY_ENC_IPV6_DST,
 951                                &mask->enc_ipv6.dst,
 952                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 953                                sizeof(key->enc_ipv6.dst));
 954         }
 955
 956         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
 957                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
 958                        sizeof(key->enc_key_id.keyid));
 959
 960         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
 961                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
 962                        sizeof(key->enc_tp.src));
 963
 964         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
 965                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
 966                        sizeof(key->enc_tp.dst));
 967
 968         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
 969
 970         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
 971                 ret = fl_set_enc_opt(tb, key, mask, extack);
 972                 if (ret)
 973                         return ret;
 974         }
 975
 976         if (tb[TCA_FLOWER_KEY_FLAGS])
 977                 ret = fl_set_key_flags(tb, &key->control.flags, &mask->control.flags);
 978
 979         return ret;
 980 }
 981
 982 static void fl_mask_copy(struct fl_flow_mask *dst,
 983                          struct fl_flow_mask *src)
 984 {
 985         const void *psrc = fl_key_get_start(&src->key, src);
 986         void *pdst = fl_key_get_start(&dst->key, src);
 987
 988         memcpy(pdst, psrc, fl_mask_range(src));
 989         dst->range = src->range;
 990 }
 991
 992 static const struct rhashtable_params fl_ht_params = {
 993         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
 994         .head_offset = offsetof(struct cls_fl_filter, ht_node),
 995         .automatic_shrinking = true,
 996 };
 997
 998 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
 999 {
1000         mask->filter_ht_params = fl_ht_params;
1001         mask->filter_ht_params.key_len = fl_mask_range(mask);
1002         mask->filter_ht_params.key_offset += mask->range.start;
1003
1004         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1005 }
1006
1007 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1008 #define FL_KEY_MEMBER_SIZE(member) FIELD_SIZEOF(struct fl_flow_key, member)
1009
1010 #define FL_KEY_IS_MASKED(mask, member)                                          \
1011         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1012                    0, FL_KEY_MEMBER_SIZE(member))                               \
1013
1014 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1015         do {                                                                    \
1016                 keys[cnt].key_id = id;                                          \
1017                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1018                 cnt++;                                                          \
1019         } while(0);
1020
1021 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1022         do {                                                                    \
1023                 if (FL_KEY_IS_MASKED(mask, member))                             \
1024                         FL_KEY_SET(keys, cnt, id, member);                      \
1025         } while(0);
1026
1027 static void fl_init_dissector(struct flow_dissector *dissector,
1028                               struct fl_flow_key *mask)
1029 {
1030         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1031         size_t cnt = 0;
1032
1033         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1034         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1035         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1036                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1037         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1038                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1039         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1040                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1041         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1042                              FLOW_DISSECTOR_KEY_PORTS, tp);
1043         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1044                              FLOW_DISSECTOR_KEY_IP, ip);
1045         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1046                              FLOW_DISSECTOR_KEY_TCP, tcp);
1047         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1048                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1049         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1050                              FLOW_DISSECTOR_KEY_ARP, arp);
1051         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1052                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1053         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1054                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1055         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1056                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1057         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1058                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1059         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1060                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1061         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1062                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1063         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1064             FL_KEY_IS_MASKED(mask, enc_ipv6))
1065                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1066                            enc_control);
1067         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1068                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1069         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1070                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1071         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1072                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1073
1074         skb_flow_dissector_init(dissector, keys, cnt);
1075 }
1076
1077 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1078                                                struct fl_flow_mask *mask)
1079 {
1080         struct fl_flow_mask *newmask;
1081         int err;
1082
1083         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1084         if (!newmask)
1085                 return ERR_PTR(-ENOMEM);
1086
1087         fl_mask_copy(newmask, mask);
1088
1089         err = fl_init_mask_hashtable(newmask);
1090         if (err)
1091                 goto errout_free;
1092
1093         fl_init_dissector(&newmask->dissector, &newmask->key);
1094
1095         INIT_LIST_HEAD_RCU(&newmask->filters);
1096
1097         err = rhashtable_insert_fast(&head->ht, &newmask->ht_node,
1098                                      mask_ht_params);
1099         if (err)
1100                 goto errout_destroy;
1101
1102         list_add_tail_rcu(&newmask->list, &head->masks);
1103
1104         return newmask;
1105
1106 errout_destroy:
1107         rhashtable_destroy(&newmask->ht);
1108 errout_free:
1109         kfree(newmask);
1110
1111         return ERR_PTR(err);
1112 }
1113
1114 static int fl_check_assign_mask(struct cls_fl_head *head,
1115                                 struct cls_fl_filter *fnew,
1116                                 struct cls_fl_filter *fold,
1117                                 struct fl_flow_mask *mask)
1118 {
1119         struct fl_flow_mask *newmask;
1120
1121         fnew->mask = rhashtable_lookup_fast(&head->ht, mask, mask_ht_params);
1122         if (!fnew->mask) {
1123                 if (fold)
1124                         return -EINVAL;
1125
1126                 newmask = fl_create_new_mask(head, mask);
1127                 if (IS_ERR(newmask))
1128                         return PTR_ERR(newmask);
1129
1130                 fnew->mask = newmask;
1131         } else if (fold && fold->mask != fnew->mask) {
1132                 return -EINVAL;
1133         }
1134
1135         return 0;
1136 }
1137
1138 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1139                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1140                         unsigned long base, struct nlattr **tb,
1141                         struct nlattr *est, bool ovr,
1142                         struct fl_flow_tmplt *tmplt,
1143                         struct netlink_ext_ack *extack)
1144 {
1145         int err;
1146
1147         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, extack);
1148         if (err < 0)
1149                 return err;
1150
1151         if (tb[TCA_FLOWER_CLASSID]) {
1152                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1153                 tcf_bind_filter(tp, &f->res, base);
1154         }
1155
1156         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1157         if (err)
1158                 return err;
1159
1160         fl_mask_update_range(mask);
1161         fl_set_masked_key(&f->mkey, &f->key, mask);
1162
1163         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1164                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1165                 return -EINVAL;
1166         }
1167
1168         return 0;
1169 }
1170
1171 static int fl_change(struct net *net, struct sk_buff *in_skb,
1172                      struct tcf_proto *tp, unsigned long base,
1173                      u32 handle, struct nlattr **tca,
1174                      void **arg, bool ovr, struct netlink_ext_ack *extack)
1175 {
1176         struct cls_fl_head *head = rtnl_dereference(tp->root);
1177         struct cls_fl_filter *fold = *arg;
1178         struct cls_fl_filter *fnew;
1179         struct nlattr **tb;
1180         struct fl_flow_mask mask = {};
1181         int err;
1182
1183         if (!tca[TCA_OPTIONS])
1184                 return -EINVAL;
1185
1186         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1187         if (!tb)
1188                 return -ENOBUFS;
1189
1190         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS],
1191                                fl_policy, NULL);
1192         if (err < 0)
1193                 goto errout_tb;
1194
1195         if (fold && handle && fold->handle != handle) {
1196                 err = -EINVAL;
1197                 goto errout_tb;
1198         }
1199
1200         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
1201         if (!fnew) {
1202                 err = -ENOBUFS;
1203                 goto errout_tb;
1204         }
1205
1206         err = tcf_exts_init(&fnew->exts, TCA_FLOWER_ACT, 0);
1207         if (err < 0)
1208                 goto errout;
1209
1210         if (!handle) {
1211                 handle = 1;
1212                 err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1213                                     INT_MAX, GFP_KERNEL);
1214         } else if (!fold) {
1215                 /* user specifies a handle and it doesn't exist */
1216                 err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1217                                     handle, GFP_KERNEL);
1218         }
1219         if (err)
1220                 goto errout;
1221         fnew->handle = handle;
1222
1223         if (tb[TCA_FLOWER_FLAGS]) {
1224                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
1225
1226                 if (!tc_flags_valid(fnew->flags)) {
1227                         err = -EINVAL;
1228                         goto errout_idr;
1229                 }
1230         }
1231
1232         err = fl_set_parms(net, tp, fnew, &mask, base, tb, tca[TCA_RATE], ovr,
1233                            tp->chain->tmplt_priv, extack);
1234         if (err)
1235                 goto errout_idr;
1236
1237         err = fl_check_assign_mask(head, fnew, fold, &mask);
1238         if (err)
1239                 goto errout_idr;
1240
1241         if (!tc_skip_sw(fnew->flags)) {
1242                 if (!fold && fl_lookup(fnew->mask, &fnew->mkey)) {
1243                         err = -EEXIST;
1244                         goto errout_mask;
1245                 }
1246
1247                 err = rhashtable_insert_fast(&fnew->mask->ht, &fnew->ht_node,
1248                                              fnew->mask->filter_ht_params);
1249                 if (err)
1250                         goto errout_mask;
1251         }
1252
1253         if (!tc_skip_hw(fnew->flags)) {
1254                 err = fl_hw_replace_filter(tp, fnew, extack);
1255                 if (err)
1256                         goto errout_mask;
1257         }
1258
1259         if (!tc_in_hw(fnew->flags))
1260                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1261
1262         if (fold) {
1263                 if (!tc_skip_sw(fold->flags))
1264                         rhashtable_remove_fast(&fold->mask->ht,
1265                                                &fold->ht_node,
1266                                                fold->mask->filter_ht_params);
1267                 if (!tc_skip_hw(fold->flags))
1268                         fl_hw_destroy_filter(tp, fold, NULL);
1269         }
1270
1271         *arg = fnew;
1272
1273         if (fold) {
1274                 idr_replace(&head->handle_idr, fnew, fnew->handle);
1275                 list_replace_rcu(&fold->list, &fnew->list);
1276                 tcf_unbind_filter(tp, &fold->res);
1277                 tcf_exts_get_net(&fold->exts);
1278                 tcf_queue_work(&fold->rwork, fl_destroy_filter_work);
1279         } else {
1280                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
1281         }
1282
1283         kfree(tb);
1284         return 0;
1285
1286 errout_mask:
1287         fl_mask_put(head, fnew->mask, false);
1288
1289 errout_idr:
1290         if (!fold)
1291                 idr_remove(&head->handle_idr, fnew->handle);
1292 errout:
1293         tcf_exts_destroy(&fnew->exts);
1294         kfree(fnew);
1295 errout_tb:
1296         kfree(tb);
1297         return err;
1298 }
1299
1300 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
1301                      struct netlink_ext_ack *extack)
1302 {
1303         struct cls_fl_head *head = rtnl_dereference(tp->root);
1304         struct cls_fl_filter *f = arg;
1305
1306         if (!tc_skip_sw(f->flags))
1307                 rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
1308                                        f->mask->filter_ht_params);
1309         __fl_delete(tp, f, extack);
1310         *last = list_empty(&head->masks);
1311         return 0;
1312 }
1313
1314 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg)
1315 {
1316         struct cls_fl_head *head = rtnl_dereference(tp->root);
1317         struct cls_fl_filter *f;
1318
1319         arg->count = arg->skip;
1320
1321         while ((f = idr_get_next_ul(&head->handle_idr,
1322                                     &arg->cookie)) != NULL) {
1323                 if (arg->fn(tp, f, arg) < 0) {
1324                         arg->stop = 1;
1325                         break;
1326                 }
1327                 arg->cookie = f->handle + 1;
1328                 arg->count++;
1329         }
1330 }
1331
1332 static int fl_reoffload(struct tcf_proto *tp, bool add, tc_setup_cb_t *cb,
1333                         void *cb_priv, struct netlink_ext_ack *extack)
1334 {
1335         struct cls_fl_head *head = rtnl_dereference(tp->root);
1336         struct tc_cls_flower_offload cls_flower = {};
1337         struct tcf_block *block = tp->chain->block;
1338         struct fl_flow_mask *mask;
1339         struct cls_fl_filter *f;
1340         int err;
1341
1342         list_for_each_entry(mask, &head->masks, list) {
1343                 list_for_each_entry(f, &mask->filters, list) {
1344                         if (tc_skip_hw(f->flags))
1345                                 continue;
1346
1347                         tc_cls_common_offload_init(&cls_flower.common, tp,
1348                                                    f->flags, extack);
1349                         cls_flower.command = add ?
1350                                 TC_CLSFLOWER_REPLACE : TC_CLSFLOWER_DESTROY;
1351                         cls_flower.cookie = (unsigned long)f;
1352                         cls_flower.dissector = &mask->dissector;
1353                         cls_flower.mask = &mask->key;
1354                         cls_flower.key = &f->mkey;
1355                         cls_flower.exts = &f->exts;
1356                         cls_flower.classid = f->res.classid;
1357
1358                         err = cb(TC_SETUP_CLSFLOWER, &cls_flower, cb_priv);
1359                         if (err) {
1360                                 if (add && tc_skip_sw(f->flags))
1361                                         return err;
1362                                 continue;
1363                         }
1364
1365                         tc_cls_offload_cnt_update(block, &f->in_hw_count,
1366                                                   &f->flags, add);
1367                 }
1368         }
1369
1370         return 0;
1371 }
1372
1373 static void fl_hw_create_tmplt(struct tcf_chain *chain,
1374                                struct fl_flow_tmplt *tmplt)
1375 {
1376         struct tc_cls_flower_offload cls_flower = {};
1377         struct tcf_block *block = chain->block;
1378         struct tcf_exts dummy_exts = { 0, };
1379
1380         cls_flower.common.chain_index = chain->index;
1381         cls_flower.command = TC_CLSFLOWER_TMPLT_CREATE;
1382         cls_flower.cookie = (unsigned long) tmplt;
1383         cls_flower.dissector = &tmplt->dissector;
1384         cls_flower.mask = &tmplt->mask;
1385         cls_flower.key = &tmplt->dummy_key;
1386         cls_flower.exts = &dummy_exts;
1387
1388         /* We don't care if driver (any of them) fails to handle this
1389          * call. It serves just as a hint for it.
1390          */
1391         tc_setup_cb_call(block, NULL, TC_SETUP_CLSFLOWER,
1392                          &cls_flower, false);
1393 }
1394
1395 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
1396                                 struct fl_flow_tmplt *tmplt)
1397 {
1398         struct tc_cls_flower_offload cls_flower = {};
1399         struct tcf_block *block = chain->block;
1400
1401         cls_flower.common.chain_index = chain->index;
1402         cls_flower.command = TC_CLSFLOWER_TMPLT_DESTROY;
1403         cls_flower.cookie = (unsigned long) tmplt;
1404
1405         tc_setup_cb_call(block, NULL, TC_SETUP_CLSFLOWER,
1406                          &cls_flower, false);
1407 }
1408
1409 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
1410                              struct nlattr **tca,
1411                              struct netlink_ext_ack *extack)
1412 {
1413         struct fl_flow_tmplt *tmplt;
1414         struct nlattr **tb;
1415         int err;
1416
1417         if (!tca[TCA_OPTIONS])
1418                 return ERR_PTR(-EINVAL);
1419
1420         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1421         if (!tb)
1422                 return ERR_PTR(-ENOBUFS);
1423         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS],
1424                                fl_policy, NULL);
1425         if (err)
1426                 goto errout_tb;
1427
1428         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
1429         if (!tmplt) {
1430                 err = -ENOMEM;
1431                 goto errout_tb;
1432         }
1433         tmplt->chain = chain;
1434         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
1435         if (err)
1436                 goto errout_tmplt;
1437         kfree(tb);
1438
1439         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
1440
1441         fl_hw_create_tmplt(chain, tmplt);
1442
1443         return tmplt;
1444
1445 errout_tmplt:
1446         kfree(tmplt);
1447 errout_tb:
1448         kfree(tb);
1449         return ERR_PTR(err);
1450 }
1451
1452 static void fl_tmplt_destroy(void *tmplt_priv)
1453 {
1454         struct fl_flow_tmplt *tmplt = tmplt_priv;
1455
1456         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
1457         kfree(tmplt);
1458 }
1459
1460 static int fl_dump_key_val(struct sk_buff *skb,
1461                            void *val, int val_type,
1462                            void *mask, int mask_type, int len)
1463 {
1464         int err;
1465
1466         if (!memchr_inv(mask, 0, len))
1467                 return 0;
1468         err = nla_put(skb, val_type, len, val);
1469         if (err)
1470                 return err;
1471         if (mask_type != TCA_FLOWER_UNSPEC) {
1472                 err = nla_put(skb, mask_type, len, mask);
1473                 if (err)
1474                         return err;
1475         }
1476         return 0;
1477 }
1478
1479 static int fl_dump_key_mpls(struct sk_buff *skb,
1480                             struct flow_dissector_key_mpls *mpls_key,
1481                             struct flow_dissector_key_mpls *mpls_mask)
1482 {
1483         int err;
1484
1485         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
1486                 return 0;
1487         if (mpls_mask->mpls_ttl) {
1488                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
1489                                  mpls_key->mpls_ttl);
1490                 if (err)
1491                         return err;
1492         }
1493         if (mpls_mask->mpls_tc) {
1494                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
1495                                  mpls_key->mpls_tc);
1496                 if (err)
1497                         return err;
1498         }
1499         if (mpls_mask->mpls_label) {
1500                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
1501                                   mpls_key->mpls_label);
1502                 if (err)
1503                         return err;
1504         }
1505         if (mpls_mask->mpls_bos) {
1506                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
1507                                  mpls_key->mpls_bos);
1508                 if (err)
1509                         return err;
1510         }
1511         return 0;
1512 }
1513
1514 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
1515                           struct flow_dissector_key_ip *key,
1516                           struct flow_dissector_key_ip *mask)
1517 {
1518         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1519         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1520         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1521         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1522
1523         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
1524             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
1525                 return -1;
1526
1527         return 0;
1528 }
1529
1530 static int fl_dump_key_vlan(struct sk_buff *skb,
1531                             int vlan_id_key, int vlan_prio_key,
1532                             struct flow_dissector_key_vlan *vlan_key,
1533                             struct flow_dissector_key_vlan *vlan_mask)
1534 {
1535         int err;
1536
1537         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
1538                 return 0;
1539         if (vlan_mask->vlan_id) {
1540                 err = nla_put_u16(skb, vlan_id_key,
1541                                   vlan_key->vlan_id);
1542                 if (err)
1543                         return err;
1544         }
1545         if (vlan_mask->vlan_priority) {
1546                 err = nla_put_u8(skb, vlan_prio_key,
1547                                  vlan_key->vlan_priority);
1548                 if (err)
1549                         return err;
1550         }
1551         return 0;
1552 }
1553
1554 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
1555                             u32 *flower_key, u32 *flower_mask,
1556                             u32 flower_flag_bit, u32 dissector_flag_bit)
1557 {
1558         if (dissector_mask & dissector_flag_bit) {
1559                 *flower_mask |= flower_flag_bit;
1560                 if (dissector_key & dissector_flag_bit)
1561                         *flower_key |= flower_flag_bit;
1562         }
1563 }
1564
1565 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
1566 {
1567         u32 key, mask;
1568         __be32 _key, _mask;
1569         int err;
1570
1571         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
1572                 return 0;
1573
1574         key = 0;
1575         mask = 0;
1576
1577         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
1578                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1579         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
1580                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1581                         FLOW_DIS_FIRST_FRAG);
1582
1583         _key = cpu_to_be32(key);
1584         _mask = cpu_to_be32(mask);
1585
1586         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
1587         if (err)
1588                 return err;
1589
1590         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
1591 }
1592
1593 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
1594                                   struct flow_dissector_key_enc_opts *enc_opts)
1595 {
1596         struct geneve_opt *opt;
1597         struct nlattr *nest;
1598         int opt_off = 0;
1599
1600         nest = nla_nest_start(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
1601         if (!nest)
1602                 goto nla_put_failure;
1603
1604         while (enc_opts->len > opt_off) {
1605                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
1606
1607                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
1608                                  opt->opt_class))
1609                         goto nla_put_failure;
1610                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
1611                                opt->type))
1612                         goto nla_put_failure;
1613                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
1614                             opt->length * 4, opt->opt_data))
1615                         goto nla_put_failure;
1616
1617                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
1618         }
1619         nla_nest_end(skb, nest);
1620         return 0;
1621
1622 nla_put_failure:
1623         nla_nest_cancel(skb, nest);
1624         return -EMSGSIZE;
1625 }
1626
1627 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
1628                                struct flow_dissector_key_enc_opts *enc_opts)
1629 {
1630         struct nlattr *nest;
1631         int err;
1632
1633         if (!enc_opts->len)
1634                 return 0;
1635
1636         nest = nla_nest_start(skb, enc_opt_type);
1637         if (!nest)
1638                 goto nla_put_failure;
1639
1640         switch (enc_opts->dst_opt_type) {
1641         case TUNNEL_GENEVE_OPT:
1642                 err = fl_dump_key_geneve_opt(skb, enc_opts);
1643                 if (err)
1644                         goto nla_put_failure;
1645                 break;
1646         default:
1647                 goto nla_put_failure;
1648         }
1649         nla_nest_end(skb, nest);
1650         return 0;
1651
1652 nla_put_failure:
1653         nla_nest_cancel(skb, nest);
1654         return -EMSGSIZE;
1655 }
1656
1657 static int fl_dump_key_enc_opt(struct sk_buff *skb,
1658                                struct flow_dissector_key_enc_opts *key_opts,
1659                                struct flow_dissector_key_enc_opts *msk_opts)
1660 {
1661         int err;
1662
1663         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
1664         if (err)
1665                 return err;
1666
1667         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
1668 }
1669
1670 static int fl_dump_key(struct sk_buff *skb, struct net *net,
1671                        struct fl_flow_key *key, struct fl_flow_key *mask)
1672 {
1673         if (mask->indev_ifindex) {
1674                 struct net_device *dev;
1675
1676                 dev = __dev_get_by_index(net, key->indev_ifindex);
1677                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
1678                         goto nla_put_failure;
1679         }
1680
1681         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1682                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1683                             sizeof(key->eth.dst)) ||
1684             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1685                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1686                             sizeof(key->eth.src)) ||
1687             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
1688                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
1689                             sizeof(key->basic.n_proto)))
1690                 goto nla_put_failure;
1691
1692         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
1693                 goto nla_put_failure;
1694
1695         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
1696                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
1697                 goto nla_put_failure;
1698
1699         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
1700                              TCA_FLOWER_KEY_CVLAN_PRIO,
1701                              &key->cvlan, &mask->cvlan) ||
1702             (mask->cvlan.vlan_tpid &&
1703              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1704                           key->cvlan.vlan_tpid)))
1705                 goto nla_put_failure;
1706
1707         if (mask->basic.n_proto) {
1708                 if (mask->cvlan.vlan_tpid) {
1709                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1710                                          key->basic.n_proto))
1711                                 goto nla_put_failure;
1712                 } else if (mask->vlan.vlan_tpid) {
1713                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1714                                          key->basic.n_proto))
1715                                 goto nla_put_failure;
1716                 }
1717         }
1718
1719         if ((key->basic.n_proto == htons(ETH_P_IP) ||
1720              key->basic.n_proto == htons(ETH_P_IPV6)) &&
1721             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1722                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1723                             sizeof(key->basic.ip_proto)) ||
1724             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
1725                 goto nla_put_failure;
1726
1727         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
1728             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1729                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1730                              sizeof(key->ipv4.src)) ||
1731              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1732                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1733                              sizeof(key->ipv4.dst))))
1734                 goto nla_put_failure;
1735         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
1736                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1737                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1738                                   sizeof(key->ipv6.src)) ||
1739                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1740                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1741                                   sizeof(key->ipv6.dst))))
1742                 goto nla_put_failure;
1743
1744         if (key->basic.ip_proto == IPPROTO_TCP &&
1745             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1746                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1747                              sizeof(key->tp.src)) ||
1748              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1749                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1750                              sizeof(key->tp.dst)) ||
1751              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1752                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1753                              sizeof(key->tcp.flags))))
1754                 goto nla_put_failure;
1755         else if (key->basic.ip_proto == IPPROTO_UDP &&
1756                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1757                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1758                                   sizeof(key->tp.src)) ||
1759                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1760                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1761                                   sizeof(key->tp.dst))))
1762                 goto nla_put_failure;
1763         else if (key->basic.ip_proto == IPPROTO_SCTP &&
1764                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1765                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1766                                   sizeof(key->tp.src)) ||
1767                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1768                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1769                                   sizeof(key->tp.dst))))
1770                 goto nla_put_failure;
1771         else if (key->basic.n_proto == htons(ETH_P_IP) &&
1772                  key->basic.ip_proto == IPPROTO_ICMP &&
1773                  (fl_dump_key_val(skb, &key->icmp.type,
1774                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
1775                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1776                                   sizeof(key->icmp.type)) ||
1777                   fl_dump_key_val(skb, &key->icmp.code,
1778                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
1779                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1780                                   sizeof(key->icmp.code))))
1781                 goto nla_put_failure;
1782         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1783                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
1784                  (fl_dump_key_val(skb, &key->icmp.type,
1785                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
1786                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1787                                   sizeof(key->icmp.type)) ||
1788                   fl_dump_key_val(skb, &key->icmp.code,
1789                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
1790                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1791                                   sizeof(key->icmp.code))))
1792                 goto nla_put_failure;
1793         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
1794                   key->basic.n_proto == htons(ETH_P_RARP)) &&
1795                  (fl_dump_key_val(skb, &key->arp.sip,
1796                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
1797                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
1798                                   sizeof(key->arp.sip)) ||
1799                   fl_dump_key_val(skb, &key->arp.tip,
1800                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
1801                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
1802                                   sizeof(key->arp.tip)) ||
1803                   fl_dump_key_val(skb, &key->arp.op,
1804                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
1805                                   TCA_FLOWER_KEY_ARP_OP_MASK,
1806                                   sizeof(key->arp.op)) ||
1807                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1808                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1809                                   sizeof(key->arp.sha)) ||
1810                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1811                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1812                                   sizeof(key->arp.tha))))
1813                 goto nla_put_failure;
1814
1815         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
1816             (fl_dump_key_val(skb, &key->enc_ipv4.src,
1817                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
1818                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1819                             sizeof(key->enc_ipv4.src)) ||
1820              fl_dump_key_val(skb, &key->enc_ipv4.dst,
1821                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
1822                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1823                              sizeof(key->enc_ipv4.dst))))
1824                 goto nla_put_failure;
1825         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
1826                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
1827                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
1828                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1829                             sizeof(key->enc_ipv6.src)) ||
1830                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
1831                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
1832                                  &mask->enc_ipv6.dst,
1833                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1834                             sizeof(key->enc_ipv6.dst))))
1835                 goto nla_put_failure;
1836
1837         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
1838                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
1839                             sizeof(key->enc_key_id)) ||
1840             fl_dump_key_val(skb, &key->enc_tp.src,
1841                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1842                             &mask->enc_tp.src,
1843                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1844                             sizeof(key->enc_tp.src)) ||
1845             fl_dump_key_val(skb, &key->enc_tp.dst,
1846                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1847                             &mask->enc_tp.dst,
1848                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1849                             sizeof(key->enc_tp.dst)) ||
1850             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
1851             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
1852                 goto nla_put_failure;
1853
1854         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
1855                 goto nla_put_failure;
1856
1857         return 0;
1858
1859 nla_put_failure:
1860         return -EMSGSIZE;
1861 }
1862
1863 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
1864                    struct sk_buff *skb, struct tcmsg *t)
1865 {
1866         struct cls_fl_filter *f = fh;
1867         struct nlattr *nest;
1868         struct fl_flow_key *key, *mask;
1869
1870         if (!f)
1871                 return skb->len;
1872
1873         t->tcm_handle = f->handle;
1874
1875         nest = nla_nest_start(skb, TCA_OPTIONS);
1876         if (!nest)
1877                 goto nla_put_failure;
1878
1879         if (f->res.classid &&
1880             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
1881                 goto nla_put_failure;
1882
1883         key = &f->key;
1884         mask = &f->mask->key;
1885
1886         if (fl_dump_key(skb, net, key, mask))
1887                 goto nla_put_failure;
1888
1889         if (!tc_skip_hw(f->flags))
1890                 fl_hw_update_stats(tp, f);
1891
1892         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
1893                 goto nla_put_failure;
1894
1895         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
1896                 goto nla_put_failure;
1897
1898         if (tcf_exts_dump(skb, &f->exts))
1899                 goto nla_put_failure;
1900
1901         nla_nest_end(skb, nest);
1902
1903         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
1904                 goto nla_put_failure;
1905
1906         return skb->len;
1907
1908 nla_put_failure:
1909         nla_nest_cancel(skb, nest);
1910         return -1;
1911 }
1912
1913 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
1914 {
1915         struct fl_flow_tmplt *tmplt = tmplt_priv;
1916         struct fl_flow_key *key, *mask;
1917         struct nlattr *nest;
1918
1919         nest = nla_nest_start(skb, TCA_OPTIONS);
1920         if (!nest)
1921                 goto nla_put_failure;
1922
1923         key = &tmplt->dummy_key;
1924         mask = &tmplt->mask;
1925
1926         if (fl_dump_key(skb, net, key, mask))
1927                 goto nla_put_failure;
1928
1929         nla_nest_end(skb, nest);
1930
1931         return skb->len;
1932
1933 nla_put_failure:
1934         nla_nest_cancel(skb, nest);
1935         return -EMSGSIZE;
1936 }
1937
1938 static void fl_bind_class(void *fh, u32 classid, unsigned long cl)
1939 {
1940         struct cls_fl_filter *f = fh;
1941
1942         if (f && f->res.classid == classid)
1943                 f->res.class = cl;
1944 }
1945
1946 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
1947         .kind           = "flower",
1948         .classify       = fl_classify,
1949         .init           = fl_init,
1950         .destroy        = fl_destroy,
1951         .get            = fl_get,
1952         .change         = fl_change,
1953         .delete         = fl_delete,
1954         .walk           = fl_walk,
1955         .reoffload      = fl_reoffload,
1956         .dump           = fl_dump,
1957         .bind_class     = fl_bind_class,
1958         .tmplt_create   = fl_tmplt_create,
1959         .tmplt_destroy  = fl_tmplt_destroy,
1960         .tmplt_dump     = fl_tmplt_dump,
1961         .owner          = THIS_MODULE,
1962 };
1963
1964 static int __init cls_fl_init(void)
1965 {
1966         return register_tcf_proto_ops(&cls_fl_ops);
1967 }
1968
1969 static void __exit cls_fl_exit(void)
1970 {
1971         unregister_tcf_proto_ops(&cls_fl_ops);
1972 }
1973
1974 module_init(cls_fl_init);
1975 module_exit(cls_fl_exit);
1976
1977 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
1978 MODULE_DESCRIPTION("Flower classifier");
1979 MODULE_LICENSE("GPL v2");