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, &tc);
 245
 246         if (tc_skip_sw(flags))
 247                 return err;
 248
 249         return 0;
 250 }
 251
 252 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f)
 253 {
 254         struct net_device *dev = tp->q->dev_queue->dev;
 255         struct tc_cls_flower_offload offload = {0};
 256         struct tc_to_netdev tc;
 257
 258         if (!tc_should_offload(dev, tp, 0))
 259                 return;
 260
 261         offload.command = TC_CLSFLOWER_STATS;
 262         offload.cookie = (unsigned long)f;
 263         offload.exts = &f->exts;
 264
 265         tc.type = TC_SETUP_CLSFLOWER;
 266         tc.cls_flower = &offload;
 267
 268         dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol, &tc);
 269 }
 270
 271 static bool fl_destroy(struct tcf_proto *tp, bool force)
 272 {
 273         struct cls_fl_head *head = rtnl_dereference(tp->root);
 274         struct cls_fl_filter *f, *next;
 275
 276         if (!force && !list_empty(&head->filters))
 277                 return false;
 278
 279         list_for_each_entry_safe(f, next, &head->filters, list) {
 280                 fl_hw_destroy_filter(tp, (unsigned long)f);
 281                 list_del_rcu(&f->list);
 282                 call_rcu(&f->rcu, fl_destroy_filter);
 283         }
 284         RCU_INIT_POINTER(tp->root, NULL);
 285         if (head->mask_assigned)
 286                 rhashtable_destroy(&head->ht);
 287         kfree_rcu(head, rcu);
 288         return true;
 289 }
 290
 291 static unsigned long fl_get(struct tcf_proto *tp, u32 handle)
 292 {
 293         struct cls_fl_head *head = rtnl_dereference(tp->root);
 294         struct cls_fl_filter *f;
 295
 296         list_for_each_entry(f, &head->filters, list)
 297                 if (f->handle == handle)
 298                         return (unsigned long) f;
 299         return 0;
 300 }
 301
 302 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 303         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 304         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 305         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 306                                             .len = IFNAMSIZ },
 307         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 308         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 309         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 310         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 311         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 312         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 313         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 314         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 315         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 316         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 317         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 318         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 319         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 320         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 321         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 322         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 323         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 324         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 325         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 326         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 327         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 328         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 329         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 330         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 331         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 332         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 333         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 334         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 335         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 336         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 337         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 338         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 339         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 340         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 341 };
 342
 343 static void fl_set_key_val(struct nlattr **tb,
 344                            void *val, int val_type,
 345                            void *mask, int mask_type, int len)
 346 {
 347         if (!tb[val_type])
 348                 return;
 349         memcpy(val, nla_data(tb[val_type]), len);
 350         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 351                 memset(mask, 0xff, len);
 352         else
 353                 memcpy(mask, nla_data(tb[mask_type]), len);
 354 }
 355
 356 static void fl_set_key_vlan(struct nlattr **tb,
 357                             struct flow_dissector_key_vlan *key_val,
 358                             struct flow_dissector_key_vlan *key_mask)
 359 {
 360 #define VLAN_PRIORITY_MASK      0x7
 361
 362         if (tb[TCA_FLOWER_KEY_VLAN_ID]) {
 363                 key_val->vlan_id =
 364                         nla_get_u16(tb[TCA_FLOWER_KEY_VLAN_ID]) & VLAN_VID_MASK;
 365                 key_mask->vlan_id = VLAN_VID_MASK;
 366         }
 367         if (tb[TCA_FLOWER_KEY_VLAN_PRIO]) {
 368                 key_val->vlan_priority =
 369                         nla_get_u8(tb[TCA_FLOWER_KEY_VLAN_PRIO]) &
 370                         VLAN_PRIORITY_MASK;
 371                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 372         }
 373 }
 374
 375 static int fl_set_key(struct net *net, struct nlattr **tb,
 376                       struct fl_flow_key *key, struct fl_flow_key *mask)
 377 {
 378         __be16 ethertype;
 379 #ifdef CONFIG_NET_CLS_IND
 380         if (tb[TCA_FLOWER_INDEV]) {
 381                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV]);
 382                 if (err < 0)
 383                         return err;
 384                 key->indev_ifindex = err;
 385                 mask->indev_ifindex = 0xffffffff;
 386         }
 387 #endif
 388
 389         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 390                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 391                        sizeof(key->eth.dst));
 392         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 393                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 394                        sizeof(key->eth.src));
 395
 396         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
 397                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
 398
 399                 if (ethertype == htons(ETH_P_8021Q)) {
 400                         fl_set_key_vlan(tb, &key->vlan, &mask->vlan);
 401                         fl_set_key_val(tb, &key->basic.n_proto,
 402                                        TCA_FLOWER_KEY_VLAN_ETH_TYPE,
 403                                        &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
 404                                        sizeof(key->basic.n_proto));
 405                 } else {
 406                         key->basic.n_proto = ethertype;
 407                         mask->basic.n_proto = cpu_to_be16(~0);
 408                 }
 409         }
 410
 411         if (key->basic.n_proto == htons(ETH_P_IP) ||
 412             key->basic.n_proto == htons(ETH_P_IPV6)) {
 413                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 414                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 415                                sizeof(key->basic.ip_proto));
 416         }
 417
 418         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
 419                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 420                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 421                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 422                                sizeof(key->ipv4.src));
 423                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 424                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 425                                sizeof(key->ipv4.dst));
 426         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
 427                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 428                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 429                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 430                                sizeof(key->ipv6.src));
 431                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 432                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 433                                sizeof(key->ipv6.dst));
 434         }
 435
 436         if (key->basic.ip_proto == IPPROTO_TCP) {
 437                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 438                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 439                                sizeof(key->tp.src));
 440                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 441                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 442                                sizeof(key->tp.dst));
 443         } else if (key->basic.ip_proto == IPPROTO_UDP) {
 444                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 445                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 446                                sizeof(key->tp.src));
 447                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 448                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 449                                sizeof(key->tp.dst));
 450         }
 451
 452         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
 453             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
 454                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 455                 fl_set_key_val(tb, &key->enc_ipv4.src,
 456                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
 457                                &mask->enc_ipv4.src,
 458                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 459                                sizeof(key->enc_ipv4.src));
 460                 fl_set_key_val(tb, &key->enc_ipv4.dst,
 461                                TCA_FLOWER_KEY_ENC_IPV4_DST,
 462                                &mask->enc_ipv4.dst,
 463                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 464                                sizeof(key->enc_ipv4.dst));
 465         }
 466
 467         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
 468             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
 469                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 470                 fl_set_key_val(tb, &key->enc_ipv6.src,
 471                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
 472                                &mask->enc_ipv6.src,
 473                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 474                                sizeof(key->enc_ipv6.src));
 475                 fl_set_key_val(tb, &key->enc_ipv6.dst,
 476                                TCA_FLOWER_KEY_ENC_IPV6_DST,
 477                                &mask->enc_ipv6.dst,
 478                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 479                                sizeof(key->enc_ipv6.dst));
 480         }
 481
 482         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
 483                        &mask->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
 484                        sizeof(key->enc_key_id.keyid));
 485
 486         return 0;
 487 }
 488
 489 static bool fl_mask_eq(struct fl_flow_mask *mask1,
 490                        struct fl_flow_mask *mask2)
 491 {
 492         const long *lmask1 = fl_key_get_start(&mask1->key, mask1);
 493         const long *lmask2 = fl_key_get_start(&mask2->key, mask2);
 494
 495         return !memcmp(&mask1->range, &mask2->range, sizeof(mask1->range)) &&
 496                !memcmp(lmask1, lmask2, fl_mask_range(mask1));
 497 }
 498
 499 static const struct rhashtable_params fl_ht_params = {
 500         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
 501         .head_offset = offsetof(struct cls_fl_filter, ht_node),
 502         .automatic_shrinking = true,
 503 };
 504
 505 static int fl_init_hashtable(struct cls_fl_head *head,
 506                              struct fl_flow_mask *mask)
 507 {
 508         head->ht_params = fl_ht_params;
 509         head->ht_params.key_len = fl_mask_range(mask);
 510         head->ht_params.key_offset += mask->range.start;
 511
 512         return rhashtable_init(&head->ht, &head->ht_params);
 513 }
 514
 515 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
 516 #define FL_KEY_MEMBER_SIZE(member) (sizeof(((struct fl_flow_key *) 0)->member))
 517
 518 #define FL_KEY_IS_MASKED(mask, member)                                          \
 519         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
 520                    0, FL_KEY_MEMBER_SIZE(member))                               \
 521
 522 #define FL_KEY_SET(keys, cnt, id, member)                                       \
 523         do {                                                                    \
 524                 keys[cnt].key_id = id;                                          \
 525                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
 526                 cnt++;                                                          \
 527         } while(0);
 528
 529 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
 530         do {                                                                    \
 531                 if (FL_KEY_IS_MASKED(mask, member))                             \
 532                         FL_KEY_SET(keys, cnt, id, member);                      \
 533         } while(0);
 534
 535 static void fl_init_dissector(struct cls_fl_head *head,
 536                               struct fl_flow_mask *mask)
 537 {
 538         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
 539         size_t cnt = 0;
 540
 541         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
 542         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
 543         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 544                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
 545         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 546                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
 547         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 548                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
 549         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 550                              FLOW_DISSECTOR_KEY_PORTS, tp);
 551         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 552                              FLOW_DISSECTOR_KEY_VLAN, vlan);
 553
 554         skb_flow_dissector_init(&head->dissector, keys, cnt);
 555 }
 556
 557 static int fl_check_assign_mask(struct cls_fl_head *head,
 558                                 struct fl_flow_mask *mask)
 559 {
 560         int err;
 561
 562         if (head->mask_assigned) {
 563                 if (!fl_mask_eq(&head->mask, mask))
 564                         return -EINVAL;
 565                 else
 566                         return 0;
 567         }
 568
 569         /* Mask is not assigned yet. So assign it and init hashtable
 570          * according to that.
 571          */
 572         err = fl_init_hashtable(head, mask);
 573         if (err)
 574                 return err;
 575         memcpy(&head->mask, mask, sizeof(head->mask));
 576         head->mask_assigned = true;
 577
 578         fl_init_dissector(head, mask);
 579
 580         return 0;
 581 }
 582
 583 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
 584                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
 585                         unsigned long base, struct nlattr **tb,
 586                         struct nlattr *est, bool ovr)
 587 {
 588         struct tcf_exts e;
 589         int err;
 590
 591         err = tcf_exts_init(&e, TCA_FLOWER_ACT, 0);
 592         if (err < 0)
 593                 return err;
 594         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 595         if (err < 0)
 596                 goto errout;
 597
 598         if (tb[TCA_FLOWER_CLASSID]) {
 599                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
 600                 tcf_bind_filter(tp, &f->res, base);
 601         }
 602
 603         err = fl_set_key(net, tb, &f->key, &mask->key);
 604         if (err)
 605                 goto errout;
 606
 607         fl_mask_update_range(mask);
 608         fl_set_masked_key(&f->mkey, &f->key, mask);
 609
 610         tcf_exts_change(tp, &f->exts, &e);
 611
 612         return 0;
 613 errout:
 614         tcf_exts_destroy(&e);
 615         return err;
 616 }
 617
 618 static u32 fl_grab_new_handle(struct tcf_proto *tp,
 619                               struct cls_fl_head *head)
 620 {
 621         unsigned int i = 0x80000000;
 622         u32 handle;
 623
 624         do {
 625                 if (++head->hgen == 0x7FFFFFFF)
 626                         head->hgen = 1;
 627         } while (--i > 0 && fl_get(tp, head->hgen));
 628
 629         if (unlikely(i == 0)) {
 630                 pr_err("Insufficient number of handles\n");
 631                 handle = 0;
 632         } else {
 633                 handle = head->hgen;
 634         }
 635
 636         return handle;
 637 }
 638
 639 static int fl_change(struct net *net, struct sk_buff *in_skb,
 640                      struct tcf_proto *tp, unsigned long base,
 641                      u32 handle, struct nlattr **tca,
 642                      unsigned long *arg, bool ovr)
 643 {
 644         struct cls_fl_head *head = rtnl_dereference(tp->root);
 645         struct cls_fl_filter *fold = (struct cls_fl_filter *) *arg;
 646         struct cls_fl_filter *fnew;
 647         struct nlattr *tb[TCA_FLOWER_MAX + 1];
 648         struct fl_flow_mask mask = {};
 649         int err;
 650
 651         if (!tca[TCA_OPTIONS])
 652                 return -EINVAL;
 653
 654         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS], fl_policy);
 655         if (err < 0)
 656                 return err;
 657
 658         if (fold && handle && fold->handle != handle)
 659                 return -EINVAL;
 660
 661         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
 662         if (!fnew)
 663                 return -ENOBUFS;
 664
 665         err = tcf_exts_init(&fnew->exts, TCA_FLOWER_ACT, 0);
 666         if (err < 0)
 667                 goto errout;
 668
 669         if (!handle) {
 670                 handle = fl_grab_new_handle(tp, head);
 671                 if (!handle) {
 672                         err = -EINVAL;
 673                         goto errout;
 674                 }
 675         }
 676         fnew->handle = handle;
 677
 678         if (tb[TCA_FLOWER_FLAGS]) {
 679                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
 680
 681                 if (!tc_flags_valid(fnew->flags)) {
 682                         err = -EINVAL;
 683                         goto errout;
 684                 }
 685         }
 686
 687         err = fl_set_parms(net, tp, fnew, &mask, base, tb, tca[TCA_RATE], ovr);
 688         if (err)
 689                 goto errout;
 690
 691         err = fl_check_assign_mask(head, &mask);
 692         if (err)
 693                 goto errout;
 694
 695         if (!tc_skip_sw(fnew->flags)) {
 696                 err = rhashtable_insert_fast(&head->ht, &fnew->ht_node,
 697                                              head->ht_params);
 698                 if (err)
 699                         goto errout;
 700         }
 701
 702         err = fl_hw_replace_filter(tp,
 703                                    &head->dissector,
 704                                    &mask.key,
 705                                    &fnew->key,
 706                                    &fnew->exts,
 707                                    (unsigned long)fnew,
 708                                    fnew->flags);
 709         if (err)
 710                 goto errout;
 711
 712         if (fold) {
 713                 rhashtable_remove_fast(&head->ht, &fold->ht_node,
 714                                        head->ht_params);
 715                 fl_hw_destroy_filter(tp, (unsigned long)fold);
 716         }
 717
 718         *arg = (unsigned long) fnew;
 719
 720         if (fold) {
 721                 list_replace_rcu(&fold->list, &fnew->list);
 722                 tcf_unbind_filter(tp, &fold->res);
 723                 call_rcu(&fold->rcu, fl_destroy_filter);
 724         } else {
 725                 list_add_tail_rcu(&fnew->list, &head->filters);
 726         }
 727
 728         return 0;
 729
 730 errout:
 731         tcf_exts_destroy(&fnew->exts);
 732         kfree(fnew);
 733         return err;
 734 }
 735
 736 static int fl_delete(struct tcf_proto *tp, unsigned long arg)
 737 {
 738         struct cls_fl_head *head = rtnl_dereference(tp->root);
 739         struct cls_fl_filter *f = (struct cls_fl_filter *) arg;
 740
 741         rhashtable_remove_fast(&head->ht, &f->ht_node,
 742                                head->ht_params);
 743         list_del_rcu(&f->list);
 744         fl_hw_destroy_filter(tp, (unsigned long)f);
 745         tcf_unbind_filter(tp, &f->res);
 746         call_rcu(&f->rcu, fl_destroy_filter);
 747         return 0;
 748 }
 749
 750 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 751 {
 752         struct cls_fl_head *head = rtnl_dereference(tp->root);
 753         struct cls_fl_filter *f;
 754
 755         list_for_each_entry_rcu(f, &head->filters, list) {
 756                 if (arg->count < arg->skip)
 757                         goto skip;
 758                 if (arg->fn(tp, (unsigned long) f, arg) < 0) {
 759                         arg->stop = 1;
 760                         break;
 761                 }
 762 skip:
 763                 arg->count++;
 764         }
 765 }
 766
 767 static int fl_dump_key_val(struct sk_buff *skb,
 768                            void *val, int val_type,
 769                            void *mask, int mask_type, int len)
 770 {
 771         int err;
 772
 773         if (!memchr_inv(mask, 0, len))
 774                 return 0;
 775         err = nla_put(skb, val_type, len, val);
 776         if (err)
 777                 return err;
 778         if (mask_type != TCA_FLOWER_UNSPEC) {
 779                 err = nla_put(skb, mask_type, len, mask);
 780                 if (err)
 781                         return err;
 782         }
 783         return 0;
 784 }
 785
 786 static int fl_dump_key_vlan(struct sk_buff *skb,
 787                             struct flow_dissector_key_vlan *vlan_key,
 788                             struct flow_dissector_key_vlan *vlan_mask)
 789 {
 790         int err;
 791
 792         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
 793                 return 0;
 794         if (vlan_mask->vlan_id) {
 795                 err = nla_put_u16(skb, TCA_FLOWER_KEY_VLAN_ID,
 796                                   vlan_key->vlan_id);
 797                 if (err)
 798                         return err;
 799         }
 800         if (vlan_mask->vlan_priority) {
 801                 err = nla_put_u8(skb, TCA_FLOWER_KEY_VLAN_PRIO,
 802                                  vlan_key->vlan_priority);
 803                 if (err)
 804                         return err;
 805         }
 806         return 0;
 807 }
 808
 809 static int fl_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 810                    struct sk_buff *skb, struct tcmsg *t)
 811 {
 812         struct cls_fl_head *head = rtnl_dereference(tp->root);
 813         struct cls_fl_filter *f = (struct cls_fl_filter *) fh;
 814         struct nlattr *nest;
 815         struct fl_flow_key *key, *mask;
 816
 817         if (!f)
 818                 return skb->len;
 819
 820         t->tcm_handle = f->handle;
 821
 822         nest = nla_nest_start(skb, TCA_OPTIONS);
 823         if (!nest)
 824                 goto nla_put_failure;
 825
 826         if (f->res.classid &&
 827             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
 828                 goto nla_put_failure;
 829
 830         key = &f->key;
 831         mask = &head->mask.key;
 832
 833         if (mask->indev_ifindex) {
 834                 struct net_device *dev;
 835
 836                 dev = __dev_get_by_index(net, key->indev_ifindex);
 837                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
 838                         goto nla_put_failure;
 839         }
 840
 841         fl_hw_update_stats(tp, f);
 842
 843         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 844                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 845                             sizeof(key->eth.dst)) ||
 846             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 847                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 848                             sizeof(key->eth.src)) ||
 849             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
 850                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
 851                             sizeof(key->basic.n_proto)))
 852                 goto nla_put_failure;
 853
 854         if (fl_dump_key_vlan(skb, &key->vlan, &mask->vlan))
 855                 goto nla_put_failure;
 856
 857         if ((key->basic.n_proto == htons(ETH_P_IP) ||
 858              key->basic.n_proto == htons(ETH_P_IPV6)) &&
 859             fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 860                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 861                             sizeof(key->basic.ip_proto)))
 862                 goto nla_put_failure;
 863
 864         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
 865             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 866                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 867                              sizeof(key->ipv4.src)) ||
 868              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 869                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 870                              sizeof(key->ipv4.dst))))
 871                 goto nla_put_failure;
 872         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
 873                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 874                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 875                                   sizeof(key->ipv6.src)) ||
 876                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 877                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 878                                   sizeof(key->ipv6.dst))))
 879                 goto nla_put_failure;
 880
 881         if (key->basic.ip_proto == IPPROTO_TCP &&
 882             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 883                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 884                              sizeof(key->tp.src)) ||
 885              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 886                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 887                              sizeof(key->tp.dst))))
 888                 goto nla_put_failure;
 889         else if (key->basic.ip_proto == IPPROTO_UDP &&
 890                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 891                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 892                                   sizeof(key->tp.src)) ||
 893                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 894                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 895                                   sizeof(key->tp.dst))))
 896                 goto nla_put_failure;
 897
 898         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
 899             (fl_dump_key_val(skb, &key->enc_ipv4.src,
 900                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
 901                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 902                             sizeof(key->enc_ipv4.src)) ||
 903              fl_dump_key_val(skb, &key->enc_ipv4.dst,
 904                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
 905                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 906                              sizeof(key->enc_ipv4.dst))))
 907                 goto nla_put_failure;
 908         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
 909                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
 910                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
 911                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 912                             sizeof(key->enc_ipv6.src)) ||
 913                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
 914                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
 915                                  &mask->enc_ipv6.dst,
 916                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 917                             sizeof(key->enc_ipv6.dst))))
 918                 goto nla_put_failure;
 919
 920         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
 921                             &mask->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
 922                             sizeof(key->enc_key_id)))
 923                 goto nla_put_failure;
 924
 925         nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags);
 926
 927         if (tcf_exts_dump(skb, &f->exts))
 928                 goto nla_put_failure;
 929
 930         nla_nest_end(skb, nest);
 931
 932         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
 933                 goto nla_put_failure;
 934
 935         return skb->len;
 936
 937 nla_put_failure:
 938         nla_nest_cancel(skb, nest);
 939         return -1;
 940 }
 941
 942 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
 943         .kind           = "flower",
 944         .classify       = fl_classify,
 945         .init           = fl_init,
 946         .destroy        = fl_destroy,
 947         .get            = fl_get,
 948         .change         = fl_change,
 949         .delete         = fl_delete,
 950         .walk           = fl_walk,
 951         .dump           = fl_dump,
 952         .owner          = THIS_MODULE,
 953 };
 954
 955 static int __init cls_fl_init(void)
 956 {
 957         return register_tcf_proto_ops(&cls_fl_ops);
 958 }
 959
 960 static void __exit cls_fl_exit(void)
 961 {
 962         unregister_tcf_proto_ops(&cls_fl_ops);
 963 }
 964
 965 module_init(cls_fl_init);
 966 module_exit(cls_fl_exit);
 967
 968 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
 969 MODULE_DESCRIPTION("Flower classifier");
 970 MODULE_LICENSE("GPL v2");