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
  28 #include <net/dst.h>
  29 #include <net/dst_metadata.h>
  30
  31 struct fl_flow_key {
  32         int     indev_ifindex;
  33         struct flow_dissector_key_control control;
  34         struct flow_dissector_key_control enc_control;
  35         struct flow_dissector_key_basic basic;
  36         struct flow_dissector_key_eth_addrs eth;
  37         struct flow_dissector_key_vlan vlan;
  38         union {
  39                 struct flow_dissector_key_ipv4_addrs ipv4;
  40                 struct flow_dissector_key_ipv6_addrs ipv6;
  41         };
  42         struct flow_dissector_key_ports tp;
  43         struct flow_dissector_key_icmp icmp;
  44         struct flow_dissector_key_arp arp;
  45         struct flow_dissector_key_keyid enc_key_id;
  46         union {
  47                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  48                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  49         };
  50         struct flow_dissector_key_ports enc_tp;
  51         struct flow_dissector_key_mpls mpls;
  52 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  53
  54 struct fl_flow_mask_range {
  55         unsigned short int start;
  56         unsigned short int end;
  57 };
  58
  59 struct fl_flow_mask {
  60         struct fl_flow_key key;
  61         struct fl_flow_mask_range range;
  62         struct rcu_head rcu;
  63 };
  64
  65 struct cls_fl_head {
  66         struct rhashtable ht;
  67         struct fl_flow_mask mask;
  68         struct flow_dissector dissector;
  69         u32 hgen;
  70         bool mask_assigned;
  71         struct list_head filters;
  72         struct rhashtable_params ht_params;
  73         union {
  74                 struct work_struct work;
  75                 struct rcu_head rcu;
  76         };
  77 };
  78
  79 struct cls_fl_filter {
  80         struct rhash_head ht_node;
  81         struct fl_flow_key mkey;
  82         struct tcf_exts exts;
  83         struct tcf_result res;
  84         struct fl_flow_key key;
  85         struct list_head list;
  86         u32 handle;
  87         u32 flags;
  88         struct rcu_head rcu;
  89         struct tc_to_netdev tc;
  90         struct net_device *hw_dev;
  91 };
  92
  93 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
  94 {
  95         return mask->range.end - mask->range.start;
  96 }
  97
  98 static void fl_mask_update_range(struct fl_flow_mask *mask)
  99 {
 100         const u8 *bytes = (const u8 *) &mask->key;
 101         size_t size = sizeof(mask->key);
 102         size_t i, first = 0, last = size - 1;
 103
 104         for (i = 0; i < sizeof(mask->key); i++) {
 105                 if (bytes[i]) {
 106                         if (!first && i)
 107                                 first = i;
 108                         last = i;
 109                 }
 110         }
 111         mask->range.start = rounddown(first, sizeof(long));
 112         mask->range.end = roundup(last + 1, sizeof(long));
 113 }
 114
 115 static void *fl_key_get_start(struct fl_flow_key *key,
 116                               const struct fl_flow_mask *mask)
 117 {
 118         return (u8 *) key + mask->range.start;
 119 }
 120
 121 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 122                               struct fl_flow_mask *mask)
 123 {
 124         const long *lkey = fl_key_get_start(key, mask);
 125         const long *lmask = fl_key_get_start(&mask->key, mask);
 126         long *lmkey = fl_key_get_start(mkey, mask);
 127         int i;
 128
 129         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 130                 *lmkey++ = *lkey++ & *lmask++;
 131 }
 132
 133 static void fl_clear_masked_range(struct fl_flow_key *key,
 134                                   struct fl_flow_mask *mask)
 135 {
 136         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 137 }
 138
 139 static struct cls_fl_filter *fl_lookup(struct cls_fl_head *head,
 140                                        struct fl_flow_key *mkey)
 141 {
 142         return rhashtable_lookup_fast(&head->ht,
 143                                       fl_key_get_start(mkey, &head->mask),
 144                                       head->ht_params);
 145 }
 146
 147 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 148                        struct tcf_result *res)
 149 {
 150         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 151         struct cls_fl_filter *f;
 152         struct fl_flow_key skb_key;
 153         struct fl_flow_key skb_mkey;
 154         struct ip_tunnel_info *info;
 155
 156         if (!atomic_read(&head->ht.nelems))
 157                 return -1;
 158
 159         fl_clear_masked_range(&skb_key, &head->mask);
 160
 161         info = skb_tunnel_info(skb);
 162         if (info) {
 163                 struct ip_tunnel_key *key = &info->key;
 164
 165                 switch (ip_tunnel_info_af(info)) {
 166                 case AF_INET:
 167                         skb_key.enc_control.addr_type =
 168                                 FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 169                         skb_key.enc_ipv4.src = key->u.ipv4.src;
 170                         skb_key.enc_ipv4.dst = key->u.ipv4.dst;
 171                         break;
 172                 case AF_INET6:
 173                         skb_key.enc_control.addr_type =
 174                                 FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 175                         skb_key.enc_ipv6.src = key->u.ipv6.src;
 176                         skb_key.enc_ipv6.dst = key->u.ipv6.dst;
 177                         break;
 178                 }
 179
 180                 skb_key.enc_key_id.keyid = tunnel_id_to_key32(key->tun_id);
 181                 skb_key.enc_tp.src = key->tp_src;
 182                 skb_key.enc_tp.dst = key->tp_dst;
 183         }
 184
 185         skb_key.indev_ifindex = skb->skb_iif;
 186         /* skb_flow_dissect() does not set n_proto in case an unknown protocol,
 187          * so do it rather here.
 188          */
 189         skb_key.basic.n_proto = skb->protocol;
 190         skb_flow_dissect(skb, &head->dissector, &skb_key, 0);
 191
 192         fl_set_masked_key(&skb_mkey, &skb_key, &head->mask);
 193
 194         f = fl_lookup(head, &skb_mkey);
 195         if (f && !tc_skip_sw(f->flags)) {
 196                 *res = f->res;
 197                 return tcf_exts_exec(skb, &f->exts, res);
 198         }
 199         return -1;
 200 }
 201
 202 static int fl_init(struct tcf_proto *tp)
 203 {
 204         struct cls_fl_head *head;
 205
 206         head = kzalloc(sizeof(*head), GFP_KERNEL);
 207         if (!head)
 208                 return -ENOBUFS;
 209
 210         INIT_LIST_HEAD_RCU(&head->filters);
 211         rcu_assign_pointer(tp->root, head);
 212
 213         return 0;
 214 }
 215
 216 static void fl_destroy_filter(struct rcu_head *head)
 217 {
 218         struct cls_fl_filter *f = container_of(head, struct cls_fl_filter, rcu);
 219
 220         tcf_exts_destroy(&f->exts);
 221         kfree(f);
 222 }
 223
 224 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f)
 225 {
 226         struct tc_cls_flower_offload offload = {0};
 227         struct net_device *dev = f->hw_dev;
 228         struct tc_to_netdev *tc = &f->tc;
 229
 230         if (!tc_can_offload(dev, tp))
 231                 return;
 232
 233         offload.command = TC_CLSFLOWER_DESTROY;
 234         offload.prio = tp->prio;
 235         offload.cookie = (unsigned long)f;
 236
 237         tc->type = TC_SETUP_CLSFLOWER;
 238         tc->cls_flower = &offload;
 239
 240         dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol, tc);
 241 }
 242
 243 static int fl_hw_replace_filter(struct tcf_proto *tp,
 244                                 struct flow_dissector *dissector,
 245                                 struct fl_flow_key *mask,
 246                                 struct cls_fl_filter *f)
 247 {
 248         struct net_device *dev = tp->q->dev_queue->dev;
 249         struct tc_cls_flower_offload offload = {0};
 250         struct tc_to_netdev *tc = &f->tc;
 251         int err;
 252
 253         if (!tc_can_offload(dev, tp)) {
 254                 if (tcf_exts_get_dev(dev, &f->exts, &f->hw_dev) ||
 255                     (f->hw_dev && !tc_can_offload(f->hw_dev, tp))) {
 256                         f->hw_dev = dev;
 257                         return tc_skip_sw(f->flags) ? -EINVAL : 0;
 258                 }
 259                 dev = f->hw_dev;
 260                 tc->egress_dev = true;
 261         } else {
 262                 f->hw_dev = dev;
 263         }
 264
 265         offload.command = TC_CLSFLOWER_REPLACE;
 266         offload.prio = tp->prio;
 267         offload.cookie = (unsigned long)f;
 268         offload.dissector = dissector;
 269         offload.mask = mask;
 270         offload.key = &f->mkey;
 271         offload.exts = &f->exts;
 272
 273         tc->type = TC_SETUP_CLSFLOWER;
 274         tc->cls_flower = &offload;
 275
 276         err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol,
 277                                             tc);
 278         if (!err)
 279                 f->flags |= TCA_CLS_FLAGS_IN_HW;
 280
 281         if (tc_skip_sw(f->flags))
 282                 return err;
 283         return 0;
 284 }
 285
 286 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f)
 287 {
 288         struct tc_cls_flower_offload offload = {0};
 289         struct net_device *dev = f->hw_dev;
 290         struct tc_to_netdev *tc = &f->tc;
 291
 292         if (!tc_can_offload(dev, tp))
 293                 return;
 294
 295         offload.command = TC_CLSFLOWER_STATS;
 296         offload.prio = tp->prio;
 297         offload.cookie = (unsigned long)f;
 298         offload.exts = &f->exts;
 299
 300         tc->type = TC_SETUP_CLSFLOWER;
 301         tc->cls_flower = &offload;
 302
 303         dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol, tc);
 304 }
 305
 306 static void __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f)
 307 {
 308         list_del_rcu(&f->list);
 309         if (!tc_skip_hw(f->flags))
 310                 fl_hw_destroy_filter(tp, f);
 311         tcf_unbind_filter(tp, &f->res);
 312         call_rcu(&f->rcu, fl_destroy_filter);
 313 }
 314
 315 static void fl_destroy_sleepable(struct work_struct *work)
 316 {
 317         struct cls_fl_head *head = container_of(work, struct cls_fl_head,
 318                                                 work);
 319         if (head->mask_assigned)
 320                 rhashtable_destroy(&head->ht);
 321         kfree(head);
 322         module_put(THIS_MODULE);
 323 }
 324
 325 static void fl_destroy_rcu(struct rcu_head *rcu)
 326 {
 327         struct cls_fl_head *head = container_of(rcu, struct cls_fl_head, rcu);
 328
 329         INIT_WORK(&head->work, fl_destroy_sleepable);
 330         schedule_work(&head->work);
 331 }
 332
 333 static void fl_destroy(struct tcf_proto *tp)
 334 {
 335         struct cls_fl_head *head = rtnl_dereference(tp->root);
 336         struct cls_fl_filter *f, *next;
 337
 338         list_for_each_entry_safe(f, next, &head->filters, list)
 339                 __fl_delete(tp, f);
 340
 341         __module_get(THIS_MODULE);
 342         call_rcu(&head->rcu, fl_destroy_rcu);
 343 }
 344
 345 static unsigned long fl_get(struct tcf_proto *tp, u32 handle)
 346 {
 347         struct cls_fl_head *head = rtnl_dereference(tp->root);
 348         struct cls_fl_filter *f;
 349
 350         list_for_each_entry(f, &head->filters, list)
 351                 if (f->handle == handle)
 352                         return (unsigned long) f;
 353         return 0;
 354 }
 355
 356 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 357         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 358         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 359         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 360                                             .len = IFNAMSIZ },
 361         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 362         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 363         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 364         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 365         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 366         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 367         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 368         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 369         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 370         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 371         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 372         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 373         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 374         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 375         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 376         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 377         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 378         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 379         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 380         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 381         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 382         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 383         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 384         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 385         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 386         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 387         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 388         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 389         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 390         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 391         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 392         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 393         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 394         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 395         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 396         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 397         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 398         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 399         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 400         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 401         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 402         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 403         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 404         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 405         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 406         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 407         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 408         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 409         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 410         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 411         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 412         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 413         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 414         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 415         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 416         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 417         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 418         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 419         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 420         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 421         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 422         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 423         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 424         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 425         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 426         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 427 };
 428
 429 static void fl_set_key_val(struct nlattr **tb,
 430                            void *val, int val_type,
 431                            void *mask, int mask_type, int len)
 432 {
 433         if (!tb[val_type])
 434                 return;
 435         memcpy(val, nla_data(tb[val_type]), len);
 436         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 437                 memset(mask, 0xff, len);
 438         else
 439                 memcpy(mask, nla_data(tb[mask_type]), len);
 440 }
 441
 442 static int fl_set_key_mpls(struct nlattr **tb,
 443                            struct flow_dissector_key_mpls *key_val,
 444                            struct flow_dissector_key_mpls *key_mask)
 445 {
 446         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 447                 key_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 448                 key_mask->mpls_ttl = MPLS_TTL_MASK;
 449         }
 450         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 451                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 452
 453                 if (bos & ~MPLS_BOS_MASK)
 454                         return -EINVAL;
 455                 key_val->mpls_bos = bos;
 456                 key_mask->mpls_bos = MPLS_BOS_MASK;
 457         }
 458         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 459                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 460
 461                 if (tc & ~MPLS_TC_MASK)
 462                         return -EINVAL;
 463                 key_val->mpls_tc = tc;
 464                 key_mask->mpls_tc = MPLS_TC_MASK;
 465         }
 466         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 467                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 468
 469                 if (label & ~MPLS_LABEL_MASK)
 470                         return -EINVAL;
 471                 key_val->mpls_label = label;
 472                 key_mask->mpls_label = MPLS_LABEL_MASK;
 473         }
 474         return 0;
 475 }
 476
 477 static void fl_set_key_vlan(struct nlattr **tb,
 478                             struct flow_dissector_key_vlan *key_val,
 479                             struct flow_dissector_key_vlan *key_mask)
 480 {
 481 #define VLAN_PRIORITY_MASK      0x7
 482
 483         if (tb[TCA_FLOWER_KEY_VLAN_ID]) {
 484                 key_val->vlan_id =
 485                         nla_get_u16(tb[TCA_FLOWER_KEY_VLAN_ID]) & VLAN_VID_MASK;
 486                 key_mask->vlan_id = VLAN_VID_MASK;
 487         }
 488         if (tb[TCA_FLOWER_KEY_VLAN_PRIO]) {
 489                 key_val->vlan_priority =
 490                         nla_get_u8(tb[TCA_FLOWER_KEY_VLAN_PRIO]) &
 491                         VLAN_PRIORITY_MASK;
 492                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 493         }
 494 }
 495
 496 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
 497                             u32 *dissector_key, u32 *dissector_mask,
 498                             u32 flower_flag_bit, u32 dissector_flag_bit)
 499 {
 500         if (flower_mask & flower_flag_bit) {
 501                 *dissector_mask |= dissector_flag_bit;
 502                 if (flower_key & flower_flag_bit)
 503                         *dissector_key |= dissector_flag_bit;
 504         }
 505 }
 506
 507 static int fl_set_key_flags(struct nlattr **tb,
 508                             u32 *flags_key, u32 *flags_mask)
 509 {
 510         u32 key, mask;
 511
 512         /* mask is mandatory for flags */
 513         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK])
 514                 return -EINVAL;
 515
 516         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
 517         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
 518
 519         *flags_key  = 0;
 520         *flags_mask = 0;
 521
 522         fl_set_key_flag(key, mask, flags_key, flags_mask,
 523                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
 524
 525         return 0;
 526 }
 527
 528 static int fl_set_key(struct net *net, struct nlattr **tb,
 529                       struct fl_flow_key *key, struct fl_flow_key *mask)
 530 {
 531         __be16 ethertype;
 532         int ret = 0;
 533 #ifdef CONFIG_NET_CLS_IND
 534         if (tb[TCA_FLOWER_INDEV]) {
 535                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV]);
 536                 if (err < 0)
 537                         return err;
 538                 key->indev_ifindex = err;
 539                 mask->indev_ifindex = 0xffffffff;
 540         }
 541 #endif
 542
 543         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 544                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 545                        sizeof(key->eth.dst));
 546         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 547                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 548                        sizeof(key->eth.src));
 549
 550         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
 551                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
 552
 553                 if (ethertype == htons(ETH_P_8021Q)) {
 554                         fl_set_key_vlan(tb, &key->vlan, &mask->vlan);
 555                         fl_set_key_val(tb, &key->basic.n_proto,
 556                                        TCA_FLOWER_KEY_VLAN_ETH_TYPE,
 557                                        &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
 558                                        sizeof(key->basic.n_proto));
 559                 } else {
 560                         key->basic.n_proto = ethertype;
 561                         mask->basic.n_proto = cpu_to_be16(~0);
 562                 }
 563         }
 564
 565         if (key->basic.n_proto == htons(ETH_P_IP) ||
 566             key->basic.n_proto == htons(ETH_P_IPV6)) {
 567                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 568                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 569                                sizeof(key->basic.ip_proto));
 570         }
 571
 572         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
 573                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 574                 mask->control.addr_type = ~0;
 575                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 576                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 577                                sizeof(key->ipv4.src));
 578                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 579                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 580                                sizeof(key->ipv4.dst));
 581         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
 582                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 583                 mask->control.addr_type = ~0;
 584                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 585                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 586                                sizeof(key->ipv6.src));
 587                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 588                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 589                                sizeof(key->ipv6.dst));
 590         }
 591
 592         if (key->basic.ip_proto == IPPROTO_TCP) {
 593                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 594                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 595                                sizeof(key->tp.src));
 596                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 597                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 598                                sizeof(key->tp.dst));
 599         } else if (key->basic.ip_proto == IPPROTO_UDP) {
 600                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 601                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 602                                sizeof(key->tp.src));
 603                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 604                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 605                                sizeof(key->tp.dst));
 606         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
 607                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
 608                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
 609                                sizeof(key->tp.src));
 610                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
 611                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
 612                                sizeof(key->tp.dst));
 613         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
 614                    key->basic.ip_proto == IPPROTO_ICMP) {
 615                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
 616                                &mask->icmp.type,
 617                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
 618                                sizeof(key->icmp.type));
 619                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
 620                                &mask->icmp.code,
 621                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
 622                                sizeof(key->icmp.code));
 623         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
 624                    key->basic.ip_proto == IPPROTO_ICMPV6) {
 625                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
 626                                &mask->icmp.type,
 627                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
 628                                sizeof(key->icmp.type));
 629                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
 630                                &mask->icmp.code,
 631                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
 632                                sizeof(key->icmp.code));
 633         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
 634                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
 635                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls);
 636                 if (ret)
 637                         return ret;
 638         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
 639                    key->basic.n_proto == htons(ETH_P_RARP)) {
 640                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
 641                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
 642                                sizeof(key->arp.sip));
 643                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
 644                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
 645                                sizeof(key->arp.tip));
 646                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
 647                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
 648                                sizeof(key->arp.op));
 649                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
 650                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
 651                                sizeof(key->arp.sha));
 652                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
 653                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
 654                                sizeof(key->arp.tha));
 655         }
 656
 657         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
 658             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
 659                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 660                 mask->enc_control.addr_type = ~0;
 661                 fl_set_key_val(tb, &key->enc_ipv4.src,
 662                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
 663                                &mask->enc_ipv4.src,
 664                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 665                                sizeof(key->enc_ipv4.src));
 666                 fl_set_key_val(tb, &key->enc_ipv4.dst,
 667                                TCA_FLOWER_KEY_ENC_IPV4_DST,
 668                                &mask->enc_ipv4.dst,
 669                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 670                                sizeof(key->enc_ipv4.dst));
 671         }
 672
 673         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
 674             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
 675                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 676                 mask->enc_control.addr_type = ~0;
 677                 fl_set_key_val(tb, &key->enc_ipv6.src,
 678                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
 679                                &mask->enc_ipv6.src,
 680                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 681                                sizeof(key->enc_ipv6.src));
 682                 fl_set_key_val(tb, &key->enc_ipv6.dst,
 683                                TCA_FLOWER_KEY_ENC_IPV6_DST,
 684                                &mask->enc_ipv6.dst,
 685                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 686                                sizeof(key->enc_ipv6.dst));
 687         }
 688
 689         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
 690                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
 691                        sizeof(key->enc_key_id.keyid));
 692
 693         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
 694                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
 695                        sizeof(key->enc_tp.src));
 696
 697         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
 698                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
 699                        sizeof(key->enc_tp.dst));
 700
 701         if (tb[TCA_FLOWER_KEY_FLAGS])
 702                 ret = fl_set_key_flags(tb, &key->control.flags, &mask->control.flags);
 703
 704         return ret;
 705 }
 706
 707 static bool fl_mask_eq(struct fl_flow_mask *mask1,
 708                        struct fl_flow_mask *mask2)
 709 {
 710         const long *lmask1 = fl_key_get_start(&mask1->key, mask1);
 711         const long *lmask2 = fl_key_get_start(&mask2->key, mask2);
 712
 713         return !memcmp(&mask1->range, &mask2->range, sizeof(mask1->range)) &&
 714                !memcmp(lmask1, lmask2, fl_mask_range(mask1));
 715 }
 716
 717 static const struct rhashtable_params fl_ht_params = {
 718         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
 719         .head_offset = offsetof(struct cls_fl_filter, ht_node),
 720         .automatic_shrinking = true,
 721 };
 722
 723 static int fl_init_hashtable(struct cls_fl_head *head,
 724                              struct fl_flow_mask *mask)
 725 {
 726         head->ht_params = fl_ht_params;
 727         head->ht_params.key_len = fl_mask_range(mask);
 728         head->ht_params.key_offset += mask->range.start;
 729
 730         return rhashtable_init(&head->ht, &head->ht_params);
 731 }
 732
 733 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
 734 #define FL_KEY_MEMBER_SIZE(member) (sizeof(((struct fl_flow_key *) 0)->member))
 735
 736 #define FL_KEY_IS_MASKED(mask, member)                                          \
 737         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
 738                    0, FL_KEY_MEMBER_SIZE(member))                               \
 739
 740 #define FL_KEY_SET(keys, cnt, id, member)                                       \
 741         do {                                                                    \
 742                 keys[cnt].key_id = id;                                          \
 743                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
 744                 cnt++;                                                          \
 745         } while(0);
 746
 747 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
 748         do {                                                                    \
 749                 if (FL_KEY_IS_MASKED(mask, member))                             \
 750                         FL_KEY_SET(keys, cnt, id, member);                      \
 751         } while(0);
 752
 753 static void fl_init_dissector(struct cls_fl_head *head,
 754                               struct fl_flow_mask *mask)
 755 {
 756         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
 757         size_t cnt = 0;
 758
 759         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
 760         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
 761         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 762                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
 763         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 764                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
 765         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 766                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
 767         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 768                              FLOW_DISSECTOR_KEY_PORTS, tp);
 769         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 770                              FLOW_DISSECTOR_KEY_ICMP, icmp);
 771         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 772                              FLOW_DISSECTOR_KEY_ARP, arp);
 773         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 774                              FLOW_DISSECTOR_KEY_MPLS, mpls);
 775         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 776                              FLOW_DISSECTOR_KEY_VLAN, vlan);
 777         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 778                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
 779         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 780                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
 781         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 782                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
 783         if (FL_KEY_IS_MASKED(&mask->key, enc_ipv4) ||
 784             FL_KEY_IS_MASKED(&mask->key, enc_ipv6))
 785                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
 786                            enc_control);
 787         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 788                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
 789
 790         skb_flow_dissector_init(&head->dissector, keys, cnt);
 791 }
 792
 793 static int fl_check_assign_mask(struct cls_fl_head *head,
 794                                 struct fl_flow_mask *mask)
 795 {
 796         int err;
 797
 798         if (head->mask_assigned) {
 799                 if (!fl_mask_eq(&head->mask, mask))
 800                         return -EINVAL;
 801                 else
 802                         return 0;
 803         }
 804
 805         /* Mask is not assigned yet. So assign it and init hashtable
 806          * according to that.
 807          */
 808         err = fl_init_hashtable(head, mask);
 809         if (err)
 810                 return err;
 811         memcpy(&head->mask, mask, sizeof(head->mask));
 812         head->mask_assigned = true;
 813
 814         fl_init_dissector(head, mask);
 815
 816         return 0;
 817 }
 818
 819 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
 820                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
 821                         unsigned long base, struct nlattr **tb,
 822                         struct nlattr *est, bool ovr)
 823 {
 824         struct tcf_exts e;
 825         int err;
 826
 827         err = tcf_exts_init(&e, TCA_FLOWER_ACT, 0);
 828         if (err < 0)
 829                 return err;
 830         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 831         if (err < 0)
 832                 goto errout;
 833
 834         if (tb[TCA_FLOWER_CLASSID]) {
 835                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
 836                 tcf_bind_filter(tp, &f->res, base);
 837         }
 838
 839         err = fl_set_key(net, tb, &f->key, &mask->key);
 840         if (err)
 841                 goto errout;
 842
 843         fl_mask_update_range(mask);
 844         fl_set_masked_key(&f->mkey, &f->key, mask);
 845
 846         tcf_exts_change(tp, &f->exts, &e);
 847
 848         return 0;
 849 errout:
 850         tcf_exts_destroy(&e);
 851         return err;
 852 }
 853
 854 static u32 fl_grab_new_handle(struct tcf_proto *tp,
 855                               struct cls_fl_head *head)
 856 {
 857         unsigned int i = 0x80000000;
 858         u32 handle;
 859
 860         do {
 861                 if (++head->hgen == 0x7FFFFFFF)
 862                         head->hgen = 1;
 863         } while (--i > 0 && fl_get(tp, head->hgen));
 864
 865         if (unlikely(i == 0)) {
 866                 pr_err("Insufficient number of handles\n");
 867                 handle = 0;
 868         } else {
 869                 handle = head->hgen;
 870         }
 871
 872         return handle;
 873 }
 874
 875 static int fl_change(struct net *net, struct sk_buff *in_skb,
 876                      struct tcf_proto *tp, unsigned long base,
 877                      u32 handle, struct nlattr **tca,
 878                      unsigned long *arg, bool ovr)
 879 {
 880         struct cls_fl_head *head = rtnl_dereference(tp->root);
 881         struct cls_fl_filter *fold = (struct cls_fl_filter *) *arg;
 882         struct cls_fl_filter *fnew;
 883         struct nlattr **tb;
 884         struct fl_flow_mask mask = {};
 885         int err;
 886
 887         if (!tca[TCA_OPTIONS])
 888                 return -EINVAL;
 889
 890         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
 891         if (!tb)
 892                 return -ENOBUFS;
 893
 894         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS],
 895                                fl_policy, NULL);
 896         if (err < 0)
 897                 goto errout_tb;
 898
 899         if (fold && handle && fold->handle != handle) {
 900                 err = -EINVAL;
 901                 goto errout_tb;
 902         }
 903
 904         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
 905         if (!fnew) {
 906                 err = -ENOBUFS;
 907                 goto errout_tb;
 908         }
 909
 910         err = tcf_exts_init(&fnew->exts, TCA_FLOWER_ACT, 0);
 911         if (err < 0)
 912                 goto errout;
 913
 914         if (!handle) {
 915                 handle = fl_grab_new_handle(tp, head);
 916                 if (!handle) {
 917                         err = -EINVAL;
 918                         goto errout;
 919                 }
 920         }
 921         fnew->handle = handle;
 922
 923         if (tb[TCA_FLOWER_FLAGS]) {
 924                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
 925
 926                 if (!tc_flags_valid(fnew->flags)) {
 927                         err = -EINVAL;
 928                         goto errout;
 929                 }
 930         }
 931
 932         err = fl_set_parms(net, tp, fnew, &mask, base, tb, tca[TCA_RATE], ovr);
 933         if (err)
 934                 goto errout;
 935
 936         err = fl_check_assign_mask(head, &mask);
 937         if (err)
 938                 goto errout;
 939
 940         if (!tc_skip_sw(fnew->flags)) {
 941                 if (!fold && fl_lookup(head, &fnew->mkey)) {
 942                         err = -EEXIST;
 943                         goto errout;
 944                 }
 945
 946                 err = rhashtable_insert_fast(&head->ht, &fnew->ht_node,
 947                                              head->ht_params);
 948                 if (err)
 949                         goto errout;
 950         }
 951
 952         if (!tc_skip_hw(fnew->flags)) {
 953                 err = fl_hw_replace_filter(tp,
 954                                            &head->dissector,
 955                                            &mask.key,
 956                                            fnew);
 957                 if (err)
 958                         goto errout;
 959         }
 960
 961         if (!tc_in_hw(fnew->flags))
 962                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
 963
 964         if (fold) {
 965                 if (!tc_skip_sw(fold->flags))
 966                         rhashtable_remove_fast(&head->ht, &fold->ht_node,
 967                                                head->ht_params);
 968                 if (!tc_skip_hw(fold->flags))
 969                         fl_hw_destroy_filter(tp, fold);
 970         }
 971
 972         *arg = (unsigned long) fnew;
 973
 974         if (fold) {
 975                 list_replace_rcu(&fold->list, &fnew->list);
 976                 tcf_unbind_filter(tp, &fold->res);
 977                 call_rcu(&fold->rcu, fl_destroy_filter);
 978         } else {
 979                 list_add_tail_rcu(&fnew->list, &head->filters);
 980         }
 981
 982         kfree(tb);
 983         return 0;
 984
 985 errout:
 986         tcf_exts_destroy(&fnew->exts);
 987         kfree(fnew);
 988 errout_tb:
 989         kfree(tb);
 990         return err;
 991 }
 992
 993 static int fl_delete(struct tcf_proto *tp, unsigned long arg, bool *last)
 994 {
 995         struct cls_fl_head *head = rtnl_dereference(tp->root);
 996         struct cls_fl_filter *f = (struct cls_fl_filter *) arg;
 997
 998         if (!tc_skip_sw(f->flags))
 999                 rhashtable_remove_fast(&head->ht, &f->ht_node,
1000                                        head->ht_params);
1001         __fl_delete(tp, f);
1002         *last = list_empty(&head->filters);
1003         return 0;
1004 }
1005
1006 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg)
1007 {
1008         struct cls_fl_head *head = rtnl_dereference(tp->root);
1009         struct cls_fl_filter *f;
1010
1011         list_for_each_entry_rcu(f, &head->filters, list) {
1012                 if (arg->count < arg->skip)
1013                         goto skip;
1014                 if (arg->fn(tp, (unsigned long) f, arg) < 0) {
1015                         arg->stop = 1;
1016                         break;
1017                 }
1018 skip:
1019                 arg->count++;
1020         }
1021 }
1022
1023 static int fl_dump_key_val(struct sk_buff *skb,
1024                            void *val, int val_type,
1025                            void *mask, int mask_type, int len)
1026 {
1027         int err;
1028
1029         if (!memchr_inv(mask, 0, len))
1030                 return 0;
1031         err = nla_put(skb, val_type, len, val);
1032         if (err)
1033                 return err;
1034         if (mask_type != TCA_FLOWER_UNSPEC) {
1035                 err = nla_put(skb, mask_type, len, mask);
1036                 if (err)
1037                         return err;
1038         }
1039         return 0;
1040 }
1041
1042 static int fl_dump_key_mpls(struct sk_buff *skb,
1043                             struct flow_dissector_key_mpls *mpls_key,
1044                             struct flow_dissector_key_mpls *mpls_mask)
1045 {
1046         int err;
1047
1048         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
1049                 return 0;
1050         if (mpls_mask->mpls_ttl) {
1051                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
1052                                  mpls_key->mpls_ttl);
1053                 if (err)
1054                         return err;
1055         }
1056         if (mpls_mask->mpls_tc) {
1057                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
1058                                  mpls_key->mpls_tc);
1059                 if (err)
1060                         return err;
1061         }
1062         if (mpls_mask->mpls_label) {
1063                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
1064                                   mpls_key->mpls_label);
1065                 if (err)
1066                         return err;
1067         }
1068         if (mpls_mask->mpls_bos) {
1069                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
1070                                  mpls_key->mpls_bos);
1071                 if (err)
1072                         return err;
1073         }
1074         return 0;
1075 }
1076
1077 static int fl_dump_key_vlan(struct sk_buff *skb,
1078                             struct flow_dissector_key_vlan *vlan_key,
1079                             struct flow_dissector_key_vlan *vlan_mask)
1080 {
1081         int err;
1082
1083         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
1084                 return 0;
1085         if (vlan_mask->vlan_id) {
1086                 err = nla_put_u16(skb, TCA_FLOWER_KEY_VLAN_ID,
1087                                   vlan_key->vlan_id);
1088                 if (err)
1089                         return err;
1090         }
1091         if (vlan_mask->vlan_priority) {
1092                 err = nla_put_u8(skb, TCA_FLOWER_KEY_VLAN_PRIO,
1093                                  vlan_key->vlan_priority);
1094                 if (err)
1095                         return err;
1096         }
1097         return 0;
1098 }
1099
1100 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
1101                             u32 *flower_key, u32 *flower_mask,
1102                             u32 flower_flag_bit, u32 dissector_flag_bit)
1103 {
1104         if (dissector_mask & dissector_flag_bit) {
1105                 *flower_mask |= flower_flag_bit;
1106                 if (dissector_key & dissector_flag_bit)
1107                         *flower_key |= flower_flag_bit;
1108         }
1109 }
1110
1111 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
1112 {
1113         u32 key, mask;
1114         __be32 _key, _mask;
1115         int err;
1116
1117         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
1118                 return 0;
1119
1120         key = 0;
1121         mask = 0;
1122
1123         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
1124                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1125
1126         _key = cpu_to_be32(key);
1127         _mask = cpu_to_be32(mask);
1128
1129         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
1130         if (err)
1131                 return err;
1132
1133         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
1134 }
1135
1136 static int fl_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
1137                    struct sk_buff *skb, struct tcmsg *t)
1138 {
1139         struct cls_fl_head *head = rtnl_dereference(tp->root);
1140         struct cls_fl_filter *f = (struct cls_fl_filter *) fh;
1141         struct nlattr *nest;
1142         struct fl_flow_key *key, *mask;
1143
1144         if (!f)
1145                 return skb->len;
1146
1147         t->tcm_handle = f->handle;
1148
1149         nest = nla_nest_start(skb, TCA_OPTIONS);
1150         if (!nest)
1151                 goto nla_put_failure;
1152
1153         if (f->res.classid &&
1154             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
1155                 goto nla_put_failure;
1156
1157         key = &f->key;
1158         mask = &head->mask.key;
1159
1160         if (mask->indev_ifindex) {
1161                 struct net_device *dev;
1162
1163                 dev = __dev_get_by_index(net, key->indev_ifindex);
1164                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
1165                         goto nla_put_failure;
1166         }
1167
1168         if (!tc_skip_hw(f->flags))
1169                 fl_hw_update_stats(tp, f);
1170
1171         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1172                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1173                             sizeof(key->eth.dst)) ||
1174             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1175                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1176                             sizeof(key->eth.src)) ||
1177             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
1178                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
1179                             sizeof(key->basic.n_proto)))
1180                 goto nla_put_failure;
1181
1182         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
1183                 goto nla_put_failure;
1184
1185         if (fl_dump_key_vlan(skb, &key->vlan, &mask->vlan))
1186                 goto nla_put_failure;
1187
1188         if ((key->basic.n_proto == htons(ETH_P_IP) ||
1189              key->basic.n_proto == htons(ETH_P_IPV6)) &&
1190             fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1191                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1192                             sizeof(key->basic.ip_proto)))
1193                 goto nla_put_failure;
1194
1195         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
1196             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1197                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1198                              sizeof(key->ipv4.src)) ||
1199              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1200                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1201                              sizeof(key->ipv4.dst))))
1202                 goto nla_put_failure;
1203         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
1204                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1205                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1206                                   sizeof(key->ipv6.src)) ||
1207                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1208                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1209                                   sizeof(key->ipv6.dst))))
1210                 goto nla_put_failure;
1211
1212         if (key->basic.ip_proto == IPPROTO_TCP &&
1213             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1214                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1215                              sizeof(key->tp.src)) ||
1216              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1217                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1218                              sizeof(key->tp.dst))))
1219                 goto nla_put_failure;
1220         else if (key->basic.ip_proto == IPPROTO_UDP &&
1221                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1222                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1223                                   sizeof(key->tp.src)) ||
1224                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1225                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1226                                   sizeof(key->tp.dst))))
1227                 goto nla_put_failure;
1228         else if (key->basic.ip_proto == IPPROTO_SCTP &&
1229                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1230                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1231                                   sizeof(key->tp.src)) ||
1232                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1233                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1234                                   sizeof(key->tp.dst))))
1235                 goto nla_put_failure;
1236         else if (key->basic.n_proto == htons(ETH_P_IP) &&
1237                  key->basic.ip_proto == IPPROTO_ICMP &&
1238                  (fl_dump_key_val(skb, &key->icmp.type,
1239                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
1240                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1241                                   sizeof(key->icmp.type)) ||
1242                   fl_dump_key_val(skb, &key->icmp.code,
1243                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
1244                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1245                                   sizeof(key->icmp.code))))
1246                 goto nla_put_failure;
1247         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1248                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
1249                  (fl_dump_key_val(skb, &key->icmp.type,
1250                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
1251                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1252                                   sizeof(key->icmp.type)) ||
1253                   fl_dump_key_val(skb, &key->icmp.code,
1254                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
1255                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1256                                   sizeof(key->icmp.code))))
1257                 goto nla_put_failure;
1258         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
1259                   key->basic.n_proto == htons(ETH_P_RARP)) &&
1260                  (fl_dump_key_val(skb, &key->arp.sip,
1261                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
1262                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
1263                                   sizeof(key->arp.sip)) ||
1264                   fl_dump_key_val(skb, &key->arp.tip,
1265                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
1266                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
1267                                   sizeof(key->arp.tip)) ||
1268                   fl_dump_key_val(skb, &key->arp.op,
1269                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
1270                                   TCA_FLOWER_KEY_ARP_OP_MASK,
1271                                   sizeof(key->arp.op)) ||
1272                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1273                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1274                                   sizeof(key->arp.sha)) ||
1275                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1276                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1277                                   sizeof(key->arp.tha))))
1278                 goto nla_put_failure;
1279
1280         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
1281             (fl_dump_key_val(skb, &key->enc_ipv4.src,
1282                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
1283                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1284                             sizeof(key->enc_ipv4.src)) ||
1285              fl_dump_key_val(skb, &key->enc_ipv4.dst,
1286                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
1287                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1288                              sizeof(key->enc_ipv4.dst))))
1289                 goto nla_put_failure;
1290         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
1291                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
1292                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
1293                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1294                             sizeof(key->enc_ipv6.src)) ||
1295                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
1296                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
1297                                  &mask->enc_ipv6.dst,
1298                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1299                             sizeof(key->enc_ipv6.dst))))
1300                 goto nla_put_failure;
1301
1302         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
1303                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
1304                             sizeof(key->enc_key_id)) ||
1305             fl_dump_key_val(skb, &key->enc_tp.src,
1306                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1307                             &mask->enc_tp.src,
1308                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1309                             sizeof(key->enc_tp.src)) ||
1310             fl_dump_key_val(skb, &key->enc_tp.dst,
1311                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1312                             &mask->enc_tp.dst,
1313                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1314                             sizeof(key->enc_tp.dst)))
1315                 goto nla_put_failure;
1316
1317         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
1318                 goto nla_put_failure;
1319
1320         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
1321                 goto nla_put_failure;
1322
1323         if (tcf_exts_dump(skb, &f->exts))
1324                 goto nla_put_failure;
1325
1326         nla_nest_end(skb, nest);
1327
1328         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
1329                 goto nla_put_failure;
1330
1331         return skb->len;
1332
1333 nla_put_failure:
1334         nla_nest_cancel(skb, nest);
1335         return -1;
1336 }
1337
1338 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
1339         .kind           = "flower",
1340         .classify       = fl_classify,
1341         .init           = fl_init,
1342         .destroy        = fl_destroy,
1343         .get            = fl_get,
1344         .change         = fl_change,
1345         .delete         = fl_delete,
1346         .walk           = fl_walk,
1347         .dump           = fl_dump,
1348         .owner          = THIS_MODULE,
1349 };
1350
1351 static int __init cls_fl_init(void)
1352 {
1353         return register_tcf_proto_ops(&cls_fl_ops);
1354 }
1355
1356 static void __exit cls_fl_exit(void)
1357 {
1358         unregister_tcf_proto_ops(&cls_fl_ops);
1359 }
1360
1361 module_init(cls_fl_init);
1362 module_exit(cls_fl_exit);
1363
1364 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
1365 MODULE_DESCRIPTION("Flower classifier");
1366 MODULE_LICENSE("GPL v2");