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
  17 #include <linux/if_ether.h>
  18 #include <linux/in6.h>
  19 #include <linux/ip.h>
  20
  21 #include <net/sch_generic.h>
  22 #include <net/pkt_cls.h>
  23 #include <net/ip.h>
  24 #include <net/flow_dissector.h>
  25
  26 #include <net/dst.h>
  27 #include <net/dst_metadata.h>
  28
  29 struct fl_flow_key {
  30         int     indev_ifindex;
  31         struct flow_dissector_key_control control;
  32         struct flow_dissector_key_control enc_control;
  33         struct flow_dissector_key_basic basic;
  34         struct flow_dissector_key_eth_addrs eth;
  35         struct flow_dissector_key_vlan vlan;
  36         union {
  37                 struct flow_dissector_key_ipv4_addrs ipv4;
  38                 struct flow_dissector_key_ipv6_addrs ipv6;
  39         };
  40         struct flow_dissector_key_ports tp;
  41         struct flow_dissector_key_keyid enc_key_id;
  42         union {
  43                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  44                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  45         };
  46 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  47
  48 struct fl_flow_mask_range {
  49         unsigned short int start;
  50         unsigned short int end;
  51 };
  52
  53 struct fl_flow_mask {
  54         struct fl_flow_key key;
  55         struct fl_flow_mask_range range;
  56         struct rcu_head rcu;
  57 };
  58
  59 struct cls_fl_head {
  60         struct rhashtable ht;
  61         struct fl_flow_mask mask;
  62         struct flow_dissector dissector;
  63         u32 hgen;
  64         bool mask_assigned;
  65         struct list_head filters;
  66         struct rhashtable_params ht_params;
  67         struct rcu_head rcu;
  68 };
  69
  70 struct cls_fl_filter {
  71         struct rhash_head ht_node;
  72         struct fl_flow_key mkey;
  73         struct tcf_exts exts;
  74         struct tcf_result res;
  75         struct fl_flow_key key;
  76         struct list_head list;
  77         u32 handle;
  78         u32 flags;
  79         struct rcu_head rcu;
  80 };
  81
  82 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
  83 {
  84         return mask->range.end - mask->range.start;
  85 }
  86
  87 static void fl_mask_update_range(struct fl_flow_mask *mask)
  88 {
  89         const u8 *bytes = (const u8 *) &mask->key;
  90         size_t size = sizeof(mask->key);
  91         size_t i, first = 0, last = size - 1;
  92
  93         for (i = 0; i < sizeof(mask->key); i++) {
  94                 if (bytes[i]) {
  95                         if (!first && i)
  96                                 first = i;
  97                         last = i;
  98                 }
  99         }
 100         mask->range.start = rounddown(first, sizeof(long));
 101         mask->range.end = roundup(last + 1, sizeof(long));
 102 }
 103
 104 static void *fl_key_get_start(struct fl_flow_key *key,
 105                               const struct fl_flow_mask *mask)
 106 {
 107         return (u8 *) key + mask->range.start;
 108 }
 109
 110 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 111                               struct fl_flow_mask *mask)
 112 {
 113         const long *lkey = fl_key_get_start(key, mask);
 114         const long *lmask = fl_key_get_start(&mask->key, mask);
 115         long *lmkey = fl_key_get_start(mkey, mask);
 116         int i;
 117
 118         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 119                 *lmkey++ = *lkey++ & *lmask++;
 120 }
 121
 122 static void fl_clear_masked_range(struct fl_flow_key *key,
 123                                   struct fl_flow_mask *mask)
 124 {
 125         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 126 }
 127
 128 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 129                        struct tcf_result *res)
 130 {
 131         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 132         struct cls_fl_filter *f;
 133         struct fl_flow_key skb_key;
 134         struct fl_flow_key skb_mkey;
 135         struct ip_tunnel_info *info;
 136
 137         if (!atomic_read(&head->ht.nelems))
 138                 return -1;
 139
 140         fl_clear_masked_range(&skb_key, &head->mask);
 141
 142         info = skb_tunnel_info(skb);
 143         if (info) {
 144                 struct ip_tunnel_key *key = &info->key;
 145
 146                 switch (ip_tunnel_info_af(info)) {
 147                 case AF_INET:
 148                         skb_key.enc_ipv4.src = key->u.ipv4.src;
 149                         skb_key.enc_ipv4.dst = key->u.ipv4.dst;
 150                         break;
 151                 case AF_INET6:
 152                         skb_key.enc_ipv6.src = key->u.ipv6.src;
 153                         skb_key.enc_ipv6.dst = key->u.ipv6.dst;
 154                         break;
 155                 }
 156
 157                 skb_key.enc_key_id.keyid = tunnel_id_to_key32(key->tun_id);
 158         }
 159
 160         skb_key.indev_ifindex = skb->skb_iif;
 161         /* skb_flow_dissect() does not set n_proto in case an unknown protocol,
 162          * so do it rather here.
 163          */
 164         skb_key.basic.n_proto = skb->protocol;
 165         skb_flow_dissect(skb, &head->dissector, &skb_key, 0);
 166
 167         fl_set_masked_key(&skb_mkey, &skb_key, &head->mask);
 168
 169         f = rhashtable_lookup_fast(&head->ht,
 170                                    fl_key_get_start(&skb_mkey, &head->mask),
 171                                    head->ht_params);
 172         if (f && !tc_skip_sw(f->flags)) {
 173                 *res = f->res;
 174                 return tcf_exts_exec(skb, &f->exts, res);
 175         }
 176         return -1;
 177 }
 178
 179 static int fl_init(struct tcf_proto *tp)
 180 {
 181         struct cls_fl_head *head;
 182
 183         head = kzalloc(sizeof(*head), GFP_KERNEL);
 184         if (!head)
 185                 return -ENOBUFS;
 186
 187         INIT_LIST_HEAD_RCU(&head->filters);
 188         rcu_assign_pointer(tp->root, head);
 189
 190         return 0;
 191 }
 192
 193 static void fl_destroy_filter(struct rcu_head *head)
 194 {
 195         struct cls_fl_filter *f = container_of(head, struct cls_fl_filter, rcu);
 196
 197         tcf_exts_destroy(&f->exts);
 198         kfree(f);
 199 }
 200
 201 static void fl_hw_destroy_filter(struct tcf_proto *tp, unsigned long cookie)
 202 {
 203         struct net_device *dev = tp->q->dev_queue->dev;
 204         struct tc_cls_flower_offload offload = {0};
 205         struct tc_to_netdev tc;
 206
 207         if (!tc_should_offload(dev, tp, 0))
 208                 return;
 209
 210         offload.command = TC_CLSFLOWER_DESTROY;
 211         offload.cookie = cookie;
 212
 213         tc.type = TC_SETUP_CLSFLOWER;
 214         tc.cls_flower = &offload;
 215
 216         dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol, &tc);
 217 }
 218
 219 static int fl_hw_replace_filter(struct tcf_proto *tp,
 220                                 struct flow_dissector *dissector,
 221                                 struct fl_flow_key *mask,
 222                                 struct fl_flow_key *key,
 223                                 struct tcf_exts *actions,
 224                                 unsigned long cookie, u32 flags)
 225 {
 226         struct net_device *dev = tp->q->dev_queue->dev;
 227         struct tc_cls_flower_offload offload = {0};
 228         struct tc_to_netdev tc;
 229         int err;
 230
 231         if (!tc_should_offload(dev, tp, flags))
 232                 return tc_skip_sw(flags) ? -EINVAL : 0;
 233
 234         offload.command = TC_CLSFLOWER_REPLACE;
 235         offload.cookie = cookie;
 236         offload.dissector = dissector;
 237         offload.mask = mask;
 238         offload.key = key;
 239         offload.exts = actions;
 240
 241         tc.type = TC_SETUP_CLSFLOWER;
 242         tc.cls_flower = &offload;
 243
 244         err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol,
 245                                             &tc);
 246
 247         if (tc_skip_sw(flags))
 248                 return err;
 249
 250         return 0;
 251 }
 252
 253 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f)
 254 {
 255         struct net_device *dev = tp->q->dev_queue->dev;
 256         struct tc_cls_flower_offload offload = {0};
 257         struct tc_to_netdev tc;
 258
 259         if (!tc_should_offload(dev, tp, 0))
 260                 return;
 261
 262         offload.command = TC_CLSFLOWER_STATS;
 263         offload.cookie = (unsigned long)f;
 264         offload.exts = &f->exts;
 265
 266         tc.type = TC_SETUP_CLSFLOWER;
 267         tc.cls_flower = &offload;
 268
 269         dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol, &tc);
 270 }
 271
 272 static bool fl_destroy(struct tcf_proto *tp, bool force)
 273 {
 274         struct cls_fl_head *head = rtnl_dereference(tp->root);
 275         struct cls_fl_filter *f, *next;
 276
 277         if (!force && !list_empty(&head->filters))
 278                 return false;
 279
 280         list_for_each_entry_safe(f, next, &head->filters, list) {
 281                 fl_hw_destroy_filter(tp, (unsigned long)f);
 282                 list_del_rcu(&f->list);
 283                 call_rcu(&f->rcu, fl_destroy_filter);
 284         }
 285         RCU_INIT_POINTER(tp->root, NULL);
 286         if (head->mask_assigned)
 287                 rhashtable_destroy(&head->ht);
 288         kfree_rcu(head, rcu);
 289         return true;
 290 }
 291
 292 static unsigned long fl_get(struct tcf_proto *tp, u32 handle)
 293 {
 294         struct cls_fl_head *head = rtnl_dereference(tp->root);
 295         struct cls_fl_filter *f;
 296
 297         list_for_each_entry(f, &head->filters, list)
 298                 if (f->handle == handle)
 299                         return (unsigned long) f;
 300         return 0;
 301 }
 302
 303 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 304         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 305         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 306         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 307                                             .len = IFNAMSIZ },
 308         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 309         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 310         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 311         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 312         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 313         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 314         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 315         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 316         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 317         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 318         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 319         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 320         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 321         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 322         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 323         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 324         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 325         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 326         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 327         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 328         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 329         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 330         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 331         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 332         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 333         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 334         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 335         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 336         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 337         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 338         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 339         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 340         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 341         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 342 };
 343
 344 static void fl_set_key_val(struct nlattr **tb,
 345                            void *val, int val_type,
 346                            void *mask, int mask_type, int len)
 347 {
 348         if (!tb[val_type])
 349                 return;
 350         memcpy(val, nla_data(tb[val_type]), len);
 351         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 352                 memset(mask, 0xff, len);
 353         else
 354                 memcpy(mask, nla_data(tb[mask_type]), len);
 355 }
 356
 357 static void fl_set_key_vlan(struct nlattr **tb,
 358                             struct flow_dissector_key_vlan *key_val,
 359                             struct flow_dissector_key_vlan *key_mask)
 360 {
 361 #define VLAN_PRIORITY_MASK      0x7
 362
 363         if (tb[TCA_FLOWER_KEY_VLAN_ID]) {
 364                 key_val->vlan_id =
 365                         nla_get_u16(tb[TCA_FLOWER_KEY_VLAN_ID]) & VLAN_VID_MASK;
 366                 key_mask->vlan_id = VLAN_VID_MASK;
 367         }
 368         if (tb[TCA_FLOWER_KEY_VLAN_PRIO]) {
 369                 key_val->vlan_priority =
 370                         nla_get_u8(tb[TCA_FLOWER_KEY_VLAN_PRIO]) &
 371                         VLAN_PRIORITY_MASK;
 372                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 373         }
 374 }
 375
 376 static int fl_set_key(struct net *net, struct nlattr **tb,
 377                       struct fl_flow_key *key, struct fl_flow_key *mask)
 378 {
 379         __be16 ethertype;
 380 #ifdef CONFIG_NET_CLS_IND
 381         if (tb[TCA_FLOWER_INDEV]) {
 382                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV]);
 383                 if (err < 0)
 384                         return err;
 385                 key->indev_ifindex = err;
 386                 mask->indev_ifindex = 0xffffffff;
 387         }
 388 #endif
 389
 390         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 391                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 392                        sizeof(key->eth.dst));
 393         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 394                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 395                        sizeof(key->eth.src));
 396
 397         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
 398                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
 399
 400                 if (ethertype == htons(ETH_P_8021Q)) {
 401                         fl_set_key_vlan(tb, &key->vlan, &mask->vlan);
 402                         fl_set_key_val(tb, &key->basic.n_proto,
 403                                        TCA_FLOWER_KEY_VLAN_ETH_TYPE,
 404                                        &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
 405                                        sizeof(key->basic.n_proto));
 406                 } else {
 407                         key->basic.n_proto = ethertype;
 408                         mask->basic.n_proto = cpu_to_be16(~0);
 409                 }
 410         }
 411
 412         if (key->basic.n_proto == htons(ETH_P_IP) ||
 413             key->basic.n_proto == htons(ETH_P_IPV6)) {
 414                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 415                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 416                                sizeof(key->basic.ip_proto));
 417         }
 418
 419         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
 420                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 421                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 422                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 423                                sizeof(key->ipv4.src));
 424                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 425                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 426                                sizeof(key->ipv4.dst));
 427         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
 428                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 429                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 430                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 431                                sizeof(key->ipv6.src));
 432                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 433                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 434                                sizeof(key->ipv6.dst));
 435         }
 436
 437         if (key->basic.ip_proto == IPPROTO_TCP) {
 438                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 439                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 440                                sizeof(key->tp.src));
 441                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 442                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 443                                sizeof(key->tp.dst));
 444         } else if (key->basic.ip_proto == IPPROTO_UDP) {
 445                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 446                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 447                                sizeof(key->tp.src));
 448                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 449                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 450                                sizeof(key->tp.dst));
 451         }
 452
 453         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
 454             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
 455                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 456                 fl_set_key_val(tb, &key->enc_ipv4.src,
 457                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
 458                                &mask->enc_ipv4.src,
 459                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 460                                sizeof(key->enc_ipv4.src));
 461                 fl_set_key_val(tb, &key->enc_ipv4.dst,
 462                                TCA_FLOWER_KEY_ENC_IPV4_DST,
 463                                &mask->enc_ipv4.dst,
 464                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 465                                sizeof(key->enc_ipv4.dst));
 466         }
 467
 468         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
 469             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
 470                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 471                 fl_set_key_val(tb, &key->enc_ipv6.src,
 472                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
 473                                &mask->enc_ipv6.src,
 474                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 475                                sizeof(key->enc_ipv6.src));
 476                 fl_set_key_val(tb, &key->enc_ipv6.dst,
 477                                TCA_FLOWER_KEY_ENC_IPV6_DST,
 478                                &mask->enc_ipv6.dst,
 479                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 480                                sizeof(key->enc_ipv6.dst));
 481         }
 482
 483         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
 484                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
 485                        sizeof(key->enc_key_id.keyid));
 486
 487         return 0;
 488 }
 489
 490 static bool fl_mask_eq(struct fl_flow_mask *mask1,
 491                        struct fl_flow_mask *mask2)
 492 {
 493         const long *lmask1 = fl_key_get_start(&mask1->key, mask1);
 494         const long *lmask2 = fl_key_get_start(&mask2->key, mask2);
 495
 496         return !memcmp(&mask1->range, &mask2->range, sizeof(mask1->range)) &&
 497                !memcmp(lmask1, lmask2, fl_mask_range(mask1));
 498 }
 499
 500 static const struct rhashtable_params fl_ht_params = {
 501         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
 502         .head_offset = offsetof(struct cls_fl_filter, ht_node),
 503         .automatic_shrinking = true,
 504 };
 505
 506 static int fl_init_hashtable(struct cls_fl_head *head,
 507                              struct fl_flow_mask *mask)
 508 {
 509         head->ht_params = fl_ht_params;
 510         head->ht_params.key_len = fl_mask_range(mask);
 511         head->ht_params.key_offset += mask->range.start;
 512
 513         return rhashtable_init(&head->ht, &head->ht_params);
 514 }
 515
 516 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
 517 #define FL_KEY_MEMBER_SIZE(member) (sizeof(((struct fl_flow_key *) 0)->member))
 518
 519 #define FL_KEY_IS_MASKED(mask, member)                                          \
 520         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
 521                    0, FL_KEY_MEMBER_SIZE(member))                               \
 522
 523 #define FL_KEY_SET(keys, cnt, id, member)                                       \
 524         do {                                                                    \
 525                 keys[cnt].key_id = id;                                          \
 526                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
 527                 cnt++;                                                          \
 528         } while(0);
 529
 530 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
 531         do {                                                                    \
 532                 if (FL_KEY_IS_MASKED(mask, member))                             \
 533                         FL_KEY_SET(keys, cnt, id, member);                      \
 534         } while(0);
 535
 536 static void fl_init_dissector(struct cls_fl_head *head,
 537                               struct fl_flow_mask *mask)
 538 {
 539         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
 540         size_t cnt = 0;
 541
 542         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
 543         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
 544         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 545                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
 546         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 547                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
 548         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 549                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
 550         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 551                              FLOW_DISSECTOR_KEY_PORTS, tp);
 552         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 553                              FLOW_DISSECTOR_KEY_VLAN, vlan);
 554
 555         skb_flow_dissector_init(&head->dissector, keys, cnt);
 556 }
 557
 558 static int fl_check_assign_mask(struct cls_fl_head *head,
 559                                 struct fl_flow_mask *mask)
 560 {
 561         int err;
 562
 563         if (head->mask_assigned) {
 564                 if (!fl_mask_eq(&head->mask, mask))
 565                         return -EINVAL;
 566                 else
 567                         return 0;
 568         }
 569
 570         /* Mask is not assigned yet. So assign it and init hashtable
 571          * according to that.
 572          */
 573         err = fl_init_hashtable(head, mask);
 574         if (err)
 575                 return err;
 576         memcpy(&head->mask, mask, sizeof(head->mask));
 577         head->mask_assigned = true;
 578
 579         fl_init_dissector(head, mask);
 580
 581         return 0;
 582 }
 583
 584 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
 585                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
 586                         unsigned long base, struct nlattr **tb,
 587                         struct nlattr *est, bool ovr)
 588 {
 589         struct tcf_exts e;
 590         int err;
 591
 592         err = tcf_exts_init(&e, TCA_FLOWER_ACT, 0);
 593         if (err < 0)
 594                 return err;
 595         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 596         if (err < 0)
 597                 goto errout;
 598
 599         if (tb[TCA_FLOWER_CLASSID]) {
 600                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
 601                 tcf_bind_filter(tp, &f->res, base);
 602         }
 603
 604         err = fl_set_key(net, tb, &f->key, &mask->key);
 605         if (err)
 606                 goto errout;
 607
 608         fl_mask_update_range(mask);
 609         fl_set_masked_key(&f->mkey, &f->key, mask);
 610
 611         tcf_exts_change(tp, &f->exts, &e);
 612
 613         return 0;
 614 errout:
 615         tcf_exts_destroy(&e);
 616         return err;
 617 }
 618
 619 static u32 fl_grab_new_handle(struct tcf_proto *tp,
 620                               struct cls_fl_head *head)
 621 {
 622         unsigned int i = 0x80000000;
 623         u32 handle;
 624
 625         do {
 626                 if (++head->hgen == 0x7FFFFFFF)
 627                         head->hgen = 1;
 628         } while (--i > 0 && fl_get(tp, head->hgen));
 629
 630         if (unlikely(i == 0)) {
 631                 pr_err("Insufficient number of handles\n");
 632                 handle = 0;
 633         } else {
 634                 handle = head->hgen;
 635         }
 636
 637         return handle;
 638 }
 639
 640 static int fl_change(struct net *net, struct sk_buff *in_skb,
 641                      struct tcf_proto *tp, unsigned long base,
 642                      u32 handle, struct nlattr **tca,
 643                      unsigned long *arg, bool ovr)
 644 {
 645         struct cls_fl_head *head = rtnl_dereference(tp->root);
 646         struct cls_fl_filter *fold = (struct cls_fl_filter *) *arg;
 647         struct cls_fl_filter *fnew;
 648         struct nlattr *tb[TCA_FLOWER_MAX + 1];
 649         struct fl_flow_mask mask = {};
 650         int err;
 651
 652         if (!tca[TCA_OPTIONS])
 653                 return -EINVAL;
 654
 655         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS], fl_policy);
 656         if (err < 0)
 657                 return err;
 658
 659         if (fold && handle && fold->handle != handle)
 660                 return -EINVAL;
 661
 662         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
 663         if (!fnew)
 664                 return -ENOBUFS;
 665
 666         err = tcf_exts_init(&fnew->exts, TCA_FLOWER_ACT, 0);
 667         if (err < 0)
 668                 goto errout;
 669
 670         if (!handle) {
 671                 handle = fl_grab_new_handle(tp, head);
 672                 if (!handle) {
 673                         err = -EINVAL;
 674                         goto errout;
 675                 }
 676         }
 677         fnew->handle = handle;
 678
 679         if (tb[TCA_FLOWER_FLAGS]) {
 680                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
 681
 682                 if (!tc_flags_valid(fnew->flags)) {
 683                         err = -EINVAL;
 684                         goto errout;
 685                 }
 686         }
 687
 688         err = fl_set_parms(net, tp, fnew, &mask, base, tb, tca[TCA_RATE], ovr);
 689         if (err)
 690                 goto errout;
 691
 692         err = fl_check_assign_mask(head, &mask);
 693         if (err)
 694                 goto errout;
 695
 696         if (!tc_skip_sw(fnew->flags)) {
 697                 err = rhashtable_insert_fast(&head->ht, &fnew->ht_node,
 698                                              head->ht_params);
 699                 if (err)
 700                         goto errout;
 701         }
 702
 703         err = fl_hw_replace_filter(tp,
 704                                    &head->dissector,
 705                                    &mask.key,
 706                                    &fnew->key,
 707                                    &fnew->exts,
 708                                    (unsigned long)fnew,
 709                                    fnew->flags);
 710         if (err)
 711                 goto errout;
 712
 713         if (fold) {
 714                 rhashtable_remove_fast(&head->ht, &fold->ht_node,
 715                                        head->ht_params);
 716                 fl_hw_destroy_filter(tp, (unsigned long)fold);
 717         }
 718
 719         *arg = (unsigned long) fnew;
 720
 721         if (fold) {
 722                 list_replace_rcu(&fold->list, &fnew->list);
 723                 tcf_unbind_filter(tp, &fold->res);
 724                 call_rcu(&fold->rcu, fl_destroy_filter);
 725         } else {
 726                 list_add_tail_rcu(&fnew->list, &head->filters);
 727         }
 728
 729         return 0;
 730
 731 errout:
 732         tcf_exts_destroy(&fnew->exts);
 733         kfree(fnew);
 734         return err;
 735 }
 736
 737 static int fl_delete(struct tcf_proto *tp, unsigned long arg)
 738 {
 739         struct cls_fl_head *head = rtnl_dereference(tp->root);
 740         struct cls_fl_filter *f = (struct cls_fl_filter *) arg;
 741
 742         rhashtable_remove_fast(&head->ht, &f->ht_node,
 743                                head->ht_params);
 744         list_del_rcu(&f->list);
 745         fl_hw_destroy_filter(tp, (unsigned long)f);
 746         tcf_unbind_filter(tp, &f->res);
 747         call_rcu(&f->rcu, fl_destroy_filter);
 748         return 0;
 749 }
 750
 751 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 752 {
 753         struct cls_fl_head *head = rtnl_dereference(tp->root);
 754         struct cls_fl_filter *f;
 755
 756         list_for_each_entry_rcu(f, &head->filters, list) {
 757                 if (arg->count < arg->skip)
 758                         goto skip;
 759                 if (arg->fn(tp, (unsigned long) f, arg) < 0) {
 760                         arg->stop = 1;
 761                         break;
 762                 }
 763 skip:
 764                 arg->count++;
 765         }
 766 }
 767
 768 static int fl_dump_key_val(struct sk_buff *skb,
 769                            void *val, int val_type,
 770                            void *mask, int mask_type, int len)
 771 {
 772         int err;
 773
 774         if (!memchr_inv(mask, 0, len))
 775                 return 0;
 776         err = nla_put(skb, val_type, len, val);
 777         if (err)
 778                 return err;
 779         if (mask_type != TCA_FLOWER_UNSPEC) {
 780                 err = nla_put(skb, mask_type, len, mask);
 781                 if (err)
 782                         return err;
 783         }
 784         return 0;
 785 }
 786
 787 static int fl_dump_key_vlan(struct sk_buff *skb,
 788                             struct flow_dissector_key_vlan *vlan_key,
 789                             struct flow_dissector_key_vlan *vlan_mask)
 790 {
 791         int err;
 792
 793         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
 794                 return 0;
 795         if (vlan_mask->vlan_id) {
 796                 err = nla_put_u16(skb, TCA_FLOWER_KEY_VLAN_ID,
 797                                   vlan_key->vlan_id);
 798                 if (err)
 799                         return err;
 800         }
 801         if (vlan_mask->vlan_priority) {
 802                 err = nla_put_u8(skb, TCA_FLOWER_KEY_VLAN_PRIO,
 803                                  vlan_key->vlan_priority);
 804                 if (err)
 805                         return err;
 806         }
 807         return 0;
 808 }
 809
 810 static int fl_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 811                    struct sk_buff *skb, struct tcmsg *t)
 812 {
 813         struct cls_fl_head *head = rtnl_dereference(tp->root);
 814         struct cls_fl_filter *f = (struct cls_fl_filter *) fh;
 815         struct nlattr *nest;
 816         struct fl_flow_key *key, *mask;
 817
 818         if (!f)
 819                 return skb->len;
 820
 821         t->tcm_handle = f->handle;
 822
 823         nest = nla_nest_start(skb, TCA_OPTIONS);
 824         if (!nest)
 825                 goto nla_put_failure;
 826
 827         if (f->res.classid &&
 828             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
 829                 goto nla_put_failure;
 830
 831         key = &f->key;
 832         mask = &head->mask.key;
 833
 834         if (mask->indev_ifindex) {
 835                 struct net_device *dev;
 836
 837                 dev = __dev_get_by_index(net, key->indev_ifindex);
 838                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
 839                         goto nla_put_failure;
 840         }
 841
 842         fl_hw_update_stats(tp, f);
 843
 844         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 845                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 846                             sizeof(key->eth.dst)) ||
 847             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 848                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 849                             sizeof(key->eth.src)) ||
 850             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
 851                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
 852                             sizeof(key->basic.n_proto)))
 853                 goto nla_put_failure;
 854
 855         if (fl_dump_key_vlan(skb, &key->vlan, &mask->vlan))
 856                 goto nla_put_failure;
 857
 858         if ((key->basic.n_proto == htons(ETH_P_IP) ||
 859              key->basic.n_proto == htons(ETH_P_IPV6)) &&
 860             fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 861                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 862                             sizeof(key->basic.ip_proto)))
 863                 goto nla_put_failure;
 864
 865         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
 866             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 867                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 868                              sizeof(key->ipv4.src)) ||
 869              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 870                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 871                              sizeof(key->ipv4.dst))))
 872                 goto nla_put_failure;
 873         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
 874                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 875                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 876                                   sizeof(key->ipv6.src)) ||
 877                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 878                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 879                                   sizeof(key->ipv6.dst))))
 880                 goto nla_put_failure;
 881
 882         if (key->basic.ip_proto == IPPROTO_TCP &&
 883             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 884                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 885                              sizeof(key->tp.src)) ||
 886              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 887                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 888                              sizeof(key->tp.dst))))
 889                 goto nla_put_failure;
 890         else if (key->basic.ip_proto == IPPROTO_UDP &&
 891                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 892                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 893                                   sizeof(key->tp.src)) ||
 894                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 895                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 896                                   sizeof(key->tp.dst))))
 897                 goto nla_put_failure;
 898
 899         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
 900             (fl_dump_key_val(skb, &key->enc_ipv4.src,
 901                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
 902                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 903                             sizeof(key->enc_ipv4.src)) ||
 904              fl_dump_key_val(skb, &key->enc_ipv4.dst,
 905                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
 906                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 907                              sizeof(key->enc_ipv4.dst))))
 908                 goto nla_put_failure;
 909         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
 910                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
 911                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
 912                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 913                             sizeof(key->enc_ipv6.src)) ||
 914                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
 915                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
 916                                  &mask->enc_ipv6.dst,
 917                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 918                             sizeof(key->enc_ipv6.dst))))
 919                 goto nla_put_failure;
 920
 921         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
 922                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
 923                             sizeof(key->enc_key_id)))
 924                 goto nla_put_failure;
 925
 926         nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags);
 927
 928         if (tcf_exts_dump(skb, &f->exts))
 929                 goto nla_put_failure;
 930
 931         nla_nest_end(skb, nest);
 932
 933         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
 934                 goto nla_put_failure;
 935
 936         return skb->len;
 937
 938 nla_put_failure:
 939         nla_nest_cancel(skb, nest);
 940         return -1;
 941 }
 942
 943 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
 944         .kind           = "flower",
 945         .classify       = fl_classify,
 946         .init           = fl_init,
 947         .destroy        = fl_destroy,
 948         .get            = fl_get,
 949         .change         = fl_change,
 950         .delete         = fl_delete,
 951         .walk           = fl_walk,
 952         .dump           = fl_dump,
 953         .owner          = THIS_MODULE,
 954 };
 955
 956 static int __init cls_fl_init(void)
 957 {
 958         return register_tcf_proto_ops(&cls_fl_ops);
 959 }
 960
 961 static void __exit cls_fl_exit(void)
 962 {
 963         unregister_tcf_proto_ops(&cls_fl_ops);
 964 }
 965
 966 module_init(cls_fl_init);
 967 module_exit(cls_fl_exit);
 968
 969 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
 970 MODULE_DESCRIPTION("Flower classifier");
 971 MODULE_LICENSE("GPL v2");