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
  22 #include <net/sch_generic.h>
  23 #include <net/pkt_cls.h>
  24 #include <net/ip.h>
  25 #include <net/flow_dissector.h>
  26
  27 #include <net/dst.h>
  28 #include <net/dst_metadata.h>
  29
  30 struct fl_flow_key {
  31         int     indev_ifindex;
  32         struct flow_dissector_key_control control;
  33         struct flow_dissector_key_control enc_control;
  34         struct flow_dissector_key_basic basic;
  35         struct flow_dissector_key_eth_addrs eth;
  36         struct flow_dissector_key_vlan vlan;
  37         union {
  38                 struct flow_dissector_key_ipv4_addrs ipv4;
  39                 struct flow_dissector_key_ipv6_addrs ipv6;
  40         };
  41         struct flow_dissector_key_ports tp;
  42         struct flow_dissector_key_keyid enc_key_id;
  43         union {
  44                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  45                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  46         };
  47 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  48
  49 struct fl_flow_mask_range {
  50         unsigned short int start;
  51         unsigned short int end;
  52 };
  53
  54 struct fl_flow_mask {
  55         struct fl_flow_key key;
  56         struct fl_flow_mask_range range;
  57         struct rcu_head rcu;
  58 };
  59
  60 struct cls_fl_head {
  61         struct rhashtable ht;
  62         struct fl_flow_mask mask;
  63         struct flow_dissector dissector;
  64         u32 hgen;
  65         bool mask_assigned;
  66         struct list_head filters;
  67         struct rhashtable_params ht_params;
  68         union {
  69                 struct work_struct work;
  70                 struct rcu_head rcu;
  71         };
  72 };
  73
  74 struct cls_fl_filter {
  75         struct rhash_head ht_node;
  76         struct fl_flow_key mkey;
  77         struct tcf_exts exts;
  78         struct tcf_result res;
  79         struct fl_flow_key key;
  80         struct list_head list;
  81         u32 handle;
  82         u32 flags;
  83         struct rcu_head rcu;
  84 };
  85
  86 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
  87 {
  88         return mask->range.end - mask->range.start;
  89 }
  90
  91 static void fl_mask_update_range(struct fl_flow_mask *mask)
  92 {
  93         const u8 *bytes = (const u8 *) &mask->key;
  94         size_t size = sizeof(mask->key);
  95         size_t i, first = 0, last = size - 1;
  96
  97         for (i = 0; i < sizeof(mask->key); i++) {
  98                 if (bytes[i]) {
  99                         if (!first && i)
 100                                 first = i;
 101                         last = i;
 102                 }
 103         }
 104         mask->range.start = rounddown(first, sizeof(long));
 105         mask->range.end = roundup(last + 1, sizeof(long));
 106 }
 107
 108 static void *fl_key_get_start(struct fl_flow_key *key,
 109                               const struct fl_flow_mask *mask)
 110 {
 111         return (u8 *) key + mask->range.start;
 112 }
 113
 114 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 115                               struct fl_flow_mask *mask)
 116 {
 117         const long *lkey = fl_key_get_start(key, mask);
 118         const long *lmask = fl_key_get_start(&mask->key, mask);
 119         long *lmkey = fl_key_get_start(mkey, mask);
 120         int i;
 121
 122         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 123                 *lmkey++ = *lkey++ & *lmask++;
 124 }
 125
 126 static void fl_clear_masked_range(struct fl_flow_key *key,
 127                                   struct fl_flow_mask *mask)
 128 {
 129         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 130 }
 131
 132 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 133                        struct tcf_result *res)
 134 {
 135         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 136         struct cls_fl_filter *f;
 137         struct fl_flow_key skb_key;
 138         struct fl_flow_key skb_mkey;
 139         struct ip_tunnel_info *info;
 140
 141         if (!atomic_read(&head->ht.nelems))
 142                 return -1;
 143
 144         flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 145         fl_clear_masked_range(&skb_key, &head->mask);
 146
 147         info = skb_tunnel_info(skb);
 148         if (info) {
 149                 struct ip_tunnel_key *key = &info->key;
 150
 151                 switch (ip_tunnel_info_af(info)) {
 152                 case AF_INET:
 153                         skb_key.enc_control.addr_type =
 154                                 FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 155                         skb_key.enc_ipv4.src = key->u.ipv4.src;
 156                         skb_key.enc_ipv4.dst = key->u.ipv4.dst;
 157                         break;
 158                 case AF_INET6:
 159                         skb_key.enc_control.addr_type =
 160                                 FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 161                         skb_key.enc_ipv6.src = key->u.ipv6.src;
 162                         skb_key.enc_ipv6.dst = key->u.ipv6.dst;
 163                         break;
 164                 }
 165
 166                 skb_key.enc_key_id.keyid = tunnel_id_to_key32(key->tun_id);
 167         }
 168
 169         skb_key.indev_ifindex = skb->skb_iif;
 170         /* skb_flow_dissect() does not set n_proto in case an unknown protocol,
 171          * so do it rather here.
 172          */
 173         skb_key.basic.n_proto = skb->protocol;
 174         skb_flow_dissect(skb, &head->dissector, &skb_key, 0);
 175
 176         fl_set_masked_key(&skb_mkey, &skb_key, &head->mask);
 177
 178         f = rhashtable_lookup_fast(&head->ht,
 179                                    fl_key_get_start(&skb_mkey, &head->mask),
 180                                    head->ht_params);
 181         if (f && !tc_skip_sw(f->flags)) {
 182                 *res = f->res;
 183                 return tcf_exts_exec(skb, &f->exts, res);
 184         }
 185         return -1;
 186 }
 187
 188 static int fl_init(struct tcf_proto *tp)
 189 {
 190         struct cls_fl_head *head;
 191
 192         head = kzalloc(sizeof(*head), GFP_KERNEL);
 193         if (!head)
 194                 return -ENOBUFS;
 195
 196         INIT_LIST_HEAD_RCU(&head->filters);
 197         rcu_assign_pointer(tp->root, head);
 198
 199         return 0;
 200 }
 201
 202 static void fl_destroy_filter(struct rcu_head *head)
 203 {
 204         struct cls_fl_filter *f = container_of(head, struct cls_fl_filter, rcu);
 205
 206         tcf_exts_destroy(&f->exts);
 207         kfree(f);
 208 }
 209
 210 static void fl_hw_destroy_filter(struct tcf_proto *tp, unsigned long cookie)
 211 {
 212         struct net_device *dev = tp->q->dev_queue->dev;
 213         struct tc_cls_flower_offload offload = {0};
 214         struct tc_to_netdev tc;
 215
 216         if (!tc_should_offload(dev, tp, 0))
 217                 return;
 218
 219         offload.command = TC_CLSFLOWER_DESTROY;
 220         offload.cookie = cookie;
 221
 222         tc.type = TC_SETUP_CLSFLOWER;
 223         tc.cls_flower = &offload;
 224
 225         dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol, &tc);
 226 }
 227
 228 static int fl_hw_replace_filter(struct tcf_proto *tp,
 229                                 struct flow_dissector *dissector,
 230                                 struct fl_flow_key *mask,
 231                                 struct fl_flow_key *key,
 232                                 struct tcf_exts *actions,
 233                                 unsigned long cookie, u32 flags)
 234 {
 235         struct net_device *dev = tp->q->dev_queue->dev;
 236         struct tc_cls_flower_offload offload = {0};
 237         struct tc_to_netdev tc;
 238         int err;
 239
 240         if (!tc_should_offload(dev, tp, flags))
 241                 return tc_skip_sw(flags) ? -EINVAL : 0;
 242
 243         offload.command = TC_CLSFLOWER_REPLACE;
 244         offload.cookie = cookie;
 245         offload.dissector = dissector;
 246         offload.mask = mask;
 247         offload.key = key;
 248         offload.exts = actions;
 249
 250         tc.type = TC_SETUP_CLSFLOWER;
 251         tc.cls_flower = &offload;
 252
 253         err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol,
 254                                             &tc);
 255
 256         if (tc_skip_sw(flags))
 257                 return err;
 258
 259         return 0;
 260 }
 261
 262 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f)
 263 {
 264         struct net_device *dev = tp->q->dev_queue->dev;
 265         struct tc_cls_flower_offload offload = {0};
 266         struct tc_to_netdev tc;
 267
 268         if (!tc_should_offload(dev, tp, 0))
 269                 return;
 270
 271         offload.command = TC_CLSFLOWER_STATS;
 272         offload.cookie = (unsigned long)f;
 273         offload.exts = &f->exts;
 274
 275         tc.type = TC_SETUP_CLSFLOWER;
 276         tc.cls_flower = &offload;
 277
 278         dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, tp->protocol, &tc);
 279 }
 280
 281 static void fl_destroy_sleepable(struct work_struct *work)
 282 {
 283         struct cls_fl_head *head = container_of(work, struct cls_fl_head,
 284                                                 work);
 285         if (head->mask_assigned)
 286                 rhashtable_destroy(&head->ht);
 287         kfree(head);
 288         module_put(THIS_MODULE);
 289 }
 290
 291 static void fl_destroy_rcu(struct rcu_head *rcu)
 292 {
 293         struct cls_fl_head *head = container_of(rcu, struct cls_fl_head, rcu);
 294
 295         INIT_WORK(&head->work, fl_destroy_sleepable);
 296         schedule_work(&head->work);
 297 }
 298
 299 static bool fl_destroy(struct tcf_proto *tp, bool force)
 300 {
 301         struct cls_fl_head *head = rtnl_dereference(tp->root);
 302         struct cls_fl_filter *f, *next;
 303
 304         if (!force && !list_empty(&head->filters))
 305                 return false;
 306
 307         list_for_each_entry_safe(f, next, &head->filters, list) {
 308                 fl_hw_destroy_filter(tp, (unsigned long)f);
 309                 list_del_rcu(&f->list);
 310                 call_rcu(&f->rcu, fl_destroy_filter);
 311         }
 312
 313         __module_get(THIS_MODULE);
 314         call_rcu(&head->rcu, fl_destroy_rcu);
 315         return true;
 316 }
 317
 318 static unsigned long fl_get(struct tcf_proto *tp, u32 handle)
 319 {
 320         struct cls_fl_head *head = rtnl_dereference(tp->root);
 321         struct cls_fl_filter *f;
 322
 323         list_for_each_entry(f, &head->filters, list)
 324                 if (f->handle == handle)
 325                         return (unsigned long) f;
 326         return 0;
 327 }
 328
 329 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 330         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 331         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 332         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 333                                             .len = IFNAMSIZ },
 334         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 335         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 336         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 337         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 338         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 339         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 340         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 341         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 342         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 343         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 344         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 345         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 346         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 347         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 348         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 349         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 350         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 351         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 352         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 353         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 354         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 355         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 356         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 357         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 358         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 359         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 360         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 361         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 362         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 363         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 364         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 365         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 366         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 367         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 368         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 369 };
 370
 371 static void fl_set_key_val(struct nlattr **tb,
 372                            void *val, int val_type,
 373                            void *mask, int mask_type, int len)
 374 {
 375         if (!tb[val_type])
 376                 return;
 377         memcpy(val, nla_data(tb[val_type]), len);
 378         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 379                 memset(mask, 0xff, len);
 380         else
 381                 memcpy(mask, nla_data(tb[mask_type]), len);
 382 }
 383
 384 static void fl_set_key_vlan(struct nlattr **tb,
 385                             struct flow_dissector_key_vlan *key_val,
 386                             struct flow_dissector_key_vlan *key_mask)
 387 {
 388 #define VLAN_PRIORITY_MASK      0x7
 389
 390         if (tb[TCA_FLOWER_KEY_VLAN_ID]) {
 391                 key_val->vlan_id =
 392                         nla_get_u16(tb[TCA_FLOWER_KEY_VLAN_ID]) & VLAN_VID_MASK;
 393                 key_mask->vlan_id = VLAN_VID_MASK;
 394         }
 395         if (tb[TCA_FLOWER_KEY_VLAN_PRIO]) {
 396                 key_val->vlan_priority =
 397                         nla_get_u8(tb[TCA_FLOWER_KEY_VLAN_PRIO]) &
 398                         VLAN_PRIORITY_MASK;
 399                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 400         }
 401 }
 402
 403 static int fl_set_key(struct net *net, struct nlattr **tb,
 404                       struct fl_flow_key *key, struct fl_flow_key *mask)
 405 {
 406         __be16 ethertype;
 407 #ifdef CONFIG_NET_CLS_IND
 408         if (tb[TCA_FLOWER_INDEV]) {
 409                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV]);
 410                 if (err < 0)
 411                         return err;
 412                 key->indev_ifindex = err;
 413                 mask->indev_ifindex = 0xffffffff;
 414         }
 415 #endif
 416
 417         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 418                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 419                        sizeof(key->eth.dst));
 420         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 421                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 422                        sizeof(key->eth.src));
 423
 424         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
 425                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
 426
 427                 if (ethertype == htons(ETH_P_8021Q)) {
 428                         fl_set_key_vlan(tb, &key->vlan, &mask->vlan);
 429                         fl_set_key_val(tb, &key->basic.n_proto,
 430                                        TCA_FLOWER_KEY_VLAN_ETH_TYPE,
 431                                        &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
 432                                        sizeof(key->basic.n_proto));
 433                 } else {
 434                         key->basic.n_proto = ethertype;
 435                         mask->basic.n_proto = cpu_to_be16(~0);
 436                 }
 437         }
 438
 439         if (key->basic.n_proto == htons(ETH_P_IP) ||
 440             key->basic.n_proto == htons(ETH_P_IPV6)) {
 441                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 442                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 443                                sizeof(key->basic.ip_proto));
 444         }
 445
 446         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
 447                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 448                 mask->control.addr_type = ~0;
 449                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 450                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 451                                sizeof(key->ipv4.src));
 452                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 453                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 454                                sizeof(key->ipv4.dst));
 455         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
 456                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 457                 mask->control.addr_type = ~0;
 458                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 459                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 460                                sizeof(key->ipv6.src));
 461                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 462                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 463                                sizeof(key->ipv6.dst));
 464         }
 465
 466         if (key->basic.ip_proto == IPPROTO_TCP) {
 467                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 468                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 469                                sizeof(key->tp.src));
 470                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 471                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 472                                sizeof(key->tp.dst));
 473         } else if (key->basic.ip_proto == IPPROTO_UDP) {
 474                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 475                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 476                                sizeof(key->tp.src));
 477                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 478                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 479                                sizeof(key->tp.dst));
 480         }
 481
 482         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
 483             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
 484                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
 485                 mask->enc_control.addr_type = ~0;
 486                 fl_set_key_val(tb, &key->enc_ipv4.src,
 487                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
 488                                &mask->enc_ipv4.src,
 489                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 490                                sizeof(key->enc_ipv4.src));
 491                 fl_set_key_val(tb, &key->enc_ipv4.dst,
 492                                TCA_FLOWER_KEY_ENC_IPV4_DST,
 493                                &mask->enc_ipv4.dst,
 494                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 495                                sizeof(key->enc_ipv4.dst));
 496         }
 497
 498         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
 499             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
 500                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
 501                 mask->enc_control.addr_type = ~0;
 502                 fl_set_key_val(tb, &key->enc_ipv6.src,
 503                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
 504                                &mask->enc_ipv6.src,
 505                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 506                                sizeof(key->enc_ipv6.src));
 507                 fl_set_key_val(tb, &key->enc_ipv6.dst,
 508                                TCA_FLOWER_KEY_ENC_IPV6_DST,
 509                                &mask->enc_ipv6.dst,
 510                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 511                                sizeof(key->enc_ipv6.dst));
 512         }
 513
 514         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
 515                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
 516                        sizeof(key->enc_key_id.keyid));
 517
 518         return 0;
 519 }
 520
 521 static bool fl_mask_eq(struct fl_flow_mask *mask1,
 522                        struct fl_flow_mask *mask2)
 523 {
 524         const long *lmask1 = fl_key_get_start(&mask1->key, mask1);
 525         const long *lmask2 = fl_key_get_start(&mask2->key, mask2);
 526
 527         return !memcmp(&mask1->range, &mask2->range, sizeof(mask1->range)) &&
 528                !memcmp(lmask1, lmask2, fl_mask_range(mask1));
 529 }
 530
 531 static const struct rhashtable_params fl_ht_params = {
 532         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
 533         .head_offset = offsetof(struct cls_fl_filter, ht_node),
 534         .automatic_shrinking = true,
 535 };
 536
 537 static int fl_init_hashtable(struct cls_fl_head *head,
 538                              struct fl_flow_mask *mask)
 539 {
 540         head->ht_params = fl_ht_params;
 541         head->ht_params.key_len = fl_mask_range(mask);
 542         head->ht_params.key_offset += mask->range.start;
 543
 544         return rhashtable_init(&head->ht, &head->ht_params);
 545 }
 546
 547 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
 548 #define FL_KEY_MEMBER_SIZE(member) (sizeof(((struct fl_flow_key *) 0)->member))
 549
 550 #define FL_KEY_IS_MASKED(mask, member)                                          \
 551         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
 552                    0, FL_KEY_MEMBER_SIZE(member))                               \
 553
 554 #define FL_KEY_SET(keys, cnt, id, member)                                       \
 555         do {                                                                    \
 556                 keys[cnt].key_id = id;                                          \
 557                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
 558                 cnt++;                                                          \
 559         } while(0);
 560
 561 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
 562         do {                                                                    \
 563                 if (FL_KEY_IS_MASKED(mask, member))                             \
 564                         FL_KEY_SET(keys, cnt, id, member);                      \
 565         } while(0);
 566
 567 static void fl_init_dissector(struct cls_fl_head *head,
 568                               struct fl_flow_mask *mask)
 569 {
 570         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
 571         size_t cnt = 0;
 572
 573         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
 574         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
 575         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 576                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
 577         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 578                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
 579         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 580                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
 581         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 582                              FLOW_DISSECTOR_KEY_PORTS, tp);
 583         FL_KEY_SET_IF_MASKED(&mask->key, keys, cnt,
 584                              FLOW_DISSECTOR_KEY_VLAN, vlan);
 585
 586         skb_flow_dissector_init(&head->dissector, keys, cnt);
 587 }
 588
 589 static int fl_check_assign_mask(struct cls_fl_head *head,
 590                                 struct fl_flow_mask *mask)
 591 {
 592         int err;
 593
 594         if (head->mask_assigned) {
 595                 if (!fl_mask_eq(&head->mask, mask))
 596                         return -EINVAL;
 597                 else
 598                         return 0;
 599         }
 600
 601         /* Mask is not assigned yet. So assign it and init hashtable
 602          * according to that.
 603          */
 604         err = fl_init_hashtable(head, mask);
 605         if (err)
 606                 return err;
 607         memcpy(&head->mask, mask, sizeof(head->mask));
 608         head->mask_assigned = true;
 609
 610         fl_init_dissector(head, mask);
 611
 612         return 0;
 613 }
 614
 615 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
 616                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
 617                         unsigned long base, struct nlattr **tb,
 618                         struct nlattr *est, bool ovr)
 619 {
 620         struct tcf_exts e;
 621         int err;
 622
 623         err = tcf_exts_init(&e, TCA_FLOWER_ACT, 0);
 624         if (err < 0)
 625                 return err;
 626         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 627         if (err < 0)
 628                 goto errout;
 629
 630         if (tb[TCA_FLOWER_CLASSID]) {
 631                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
 632                 tcf_bind_filter(tp, &f->res, base);
 633         }
 634
 635         err = fl_set_key(net, tb, &f->key, &mask->key);
 636         if (err)
 637                 goto errout;
 638
 639         fl_mask_update_range(mask);
 640         fl_set_masked_key(&f->mkey, &f->key, mask);
 641
 642         tcf_exts_change(tp, &f->exts, &e);
 643
 644         return 0;
 645 errout:
 646         tcf_exts_destroy(&e);
 647         return err;
 648 }
 649
 650 static u32 fl_grab_new_handle(struct tcf_proto *tp,
 651                               struct cls_fl_head *head)
 652 {
 653         unsigned int i = 0x80000000;
 654         u32 handle;
 655
 656         do {
 657                 if (++head->hgen == 0x7FFFFFFF)
 658                         head->hgen = 1;
 659         } while (--i > 0 && fl_get(tp, head->hgen));
 660
 661         if (unlikely(i == 0)) {
 662                 pr_err("Insufficient number of handles\n");
 663                 handle = 0;
 664         } else {
 665                 handle = head->hgen;
 666         }
 667
 668         return handle;
 669 }
 670
 671 static int fl_change(struct net *net, struct sk_buff *in_skb,
 672                      struct tcf_proto *tp, unsigned long base,
 673                      u32 handle, struct nlattr **tca,
 674                      unsigned long *arg, bool ovr)
 675 {
 676         struct cls_fl_head *head = rtnl_dereference(tp->root);
 677         struct cls_fl_filter *fold = (struct cls_fl_filter *) *arg;
 678         struct cls_fl_filter *fnew;
 679         struct nlattr *tb[TCA_FLOWER_MAX + 1];
 680         struct fl_flow_mask mask = {};
 681         int err;
 682
 683         if (!tca[TCA_OPTIONS])
 684                 return -EINVAL;
 685
 686         err = nla_parse_nested(tb, TCA_FLOWER_MAX, tca[TCA_OPTIONS], fl_policy);
 687         if (err < 0)
 688                 return err;
 689
 690         if (fold && handle && fold->handle != handle)
 691                 return -EINVAL;
 692
 693         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
 694         if (!fnew)
 695                 return -ENOBUFS;
 696
 697         err = tcf_exts_init(&fnew->exts, TCA_FLOWER_ACT, 0);
 698         if (err < 0)
 699                 goto errout;
 700
 701         if (!handle) {
 702                 handle = fl_grab_new_handle(tp, head);
 703                 if (!handle) {
 704                         err = -EINVAL;
 705                         goto errout;
 706                 }
 707         }
 708         fnew->handle = handle;
 709
 710         if (tb[TCA_FLOWER_FLAGS]) {
 711                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
 712
 713                 if (!tc_flags_valid(fnew->flags)) {
 714                         err = -EINVAL;
 715                         goto errout;
 716                 }
 717         }
 718
 719         err = fl_set_parms(net, tp, fnew, &mask, base, tb, tca[TCA_RATE], ovr);
 720         if (err)
 721                 goto errout;
 722
 723         err = fl_check_assign_mask(head, &mask);
 724         if (err)
 725                 goto errout;
 726
 727         if (!tc_skip_sw(fnew->flags)) {
 728                 err = rhashtable_insert_fast(&head->ht, &fnew->ht_node,
 729                                              head->ht_params);
 730                 if (err)
 731                         goto errout;
 732         }
 733
 734         err = fl_hw_replace_filter(tp,
 735                                    &head->dissector,
 736                                    &mask.key,
 737                                    &fnew->key,
 738                                    &fnew->exts,
 739                                    (unsigned long)fnew,
 740                                    fnew->flags);
 741         if (err)
 742                 goto errout;
 743
 744         if (fold) {
 745                 if (!tc_skip_sw(fold->flags))
 746                         rhashtable_remove_fast(&head->ht, &fold->ht_node,
 747                                                head->ht_params);
 748                 fl_hw_destroy_filter(tp, (unsigned long)fold);
 749         }
 750
 751         *arg = (unsigned long) fnew;
 752
 753         if (fold) {
 754                 list_replace_rcu(&fold->list, &fnew->list);
 755                 tcf_unbind_filter(tp, &fold->res);
 756                 call_rcu(&fold->rcu, fl_destroy_filter);
 757         } else {
 758                 list_add_tail_rcu(&fnew->list, &head->filters);
 759         }
 760
 761         return 0;
 762
 763 errout:
 764         tcf_exts_destroy(&fnew->exts);
 765         kfree(fnew);
 766         return err;
 767 }
 768
 769 static int fl_delete(struct tcf_proto *tp, unsigned long arg)
 770 {
 771         struct cls_fl_head *head = rtnl_dereference(tp->root);
 772         struct cls_fl_filter *f = (struct cls_fl_filter *) arg;
 773
 774         if (!tc_skip_sw(f->flags))
 775                 rhashtable_remove_fast(&head->ht, &f->ht_node,
 776                                        head->ht_params);
 777         list_del_rcu(&f->list);
 778         fl_hw_destroy_filter(tp, (unsigned long)f);
 779         tcf_unbind_filter(tp, &f->res);
 780         call_rcu(&f->rcu, fl_destroy_filter);
 781         return 0;
 782 }
 783
 784 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 785 {
 786         struct cls_fl_head *head = rtnl_dereference(tp->root);
 787         struct cls_fl_filter *f;
 788
 789         list_for_each_entry_rcu(f, &head->filters, list) {
 790                 if (arg->count < arg->skip)
 791                         goto skip;
 792                 if (arg->fn(tp, (unsigned long) f, arg) < 0) {
 793                         arg->stop = 1;
 794                         break;
 795                 }
 796 skip:
 797                 arg->count++;
 798         }
 799 }
 800
 801 static int fl_dump_key_val(struct sk_buff *skb,
 802                            void *val, int val_type,
 803                            void *mask, int mask_type, int len)
 804 {
 805         int err;
 806
 807         if (!memchr_inv(mask, 0, len))
 808                 return 0;
 809         err = nla_put(skb, val_type, len, val);
 810         if (err)
 811                 return err;
 812         if (mask_type != TCA_FLOWER_UNSPEC) {
 813                 err = nla_put(skb, mask_type, len, mask);
 814                 if (err)
 815                         return err;
 816         }
 817         return 0;
 818 }
 819
 820 static int fl_dump_key_vlan(struct sk_buff *skb,
 821                             struct flow_dissector_key_vlan *vlan_key,
 822                             struct flow_dissector_key_vlan *vlan_mask)
 823 {
 824         int err;
 825
 826         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
 827                 return 0;
 828         if (vlan_mask->vlan_id) {
 829                 err = nla_put_u16(skb, TCA_FLOWER_KEY_VLAN_ID,
 830                                   vlan_key->vlan_id);
 831                 if (err)
 832                         return err;
 833         }
 834         if (vlan_mask->vlan_priority) {
 835                 err = nla_put_u8(skb, TCA_FLOWER_KEY_VLAN_PRIO,
 836                                  vlan_key->vlan_priority);
 837                 if (err)
 838                         return err;
 839         }
 840         return 0;
 841 }
 842
 843 static int fl_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 844                    struct sk_buff *skb, struct tcmsg *t)
 845 {
 846         struct cls_fl_head *head = rtnl_dereference(tp->root);
 847         struct cls_fl_filter *f = (struct cls_fl_filter *) fh;
 848         struct nlattr *nest;
 849         struct fl_flow_key *key, *mask;
 850
 851         if (!f)
 852                 return skb->len;
 853
 854         t->tcm_handle = f->handle;
 855
 856         nest = nla_nest_start(skb, TCA_OPTIONS);
 857         if (!nest)
 858                 goto nla_put_failure;
 859
 860         if (f->res.classid &&
 861             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
 862                 goto nla_put_failure;
 863
 864         key = &f->key;
 865         mask = &head->mask.key;
 866
 867         if (mask->indev_ifindex) {
 868                 struct net_device *dev;
 869
 870                 dev = __dev_get_by_index(net, key->indev_ifindex);
 871                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
 872                         goto nla_put_failure;
 873         }
 874
 875         fl_hw_update_stats(tp, f);
 876
 877         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
 878                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
 879                             sizeof(key->eth.dst)) ||
 880             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
 881                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
 882                             sizeof(key->eth.src)) ||
 883             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
 884                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
 885                             sizeof(key->basic.n_proto)))
 886                 goto nla_put_failure;
 887
 888         if (fl_dump_key_vlan(skb, &key->vlan, &mask->vlan))
 889                 goto nla_put_failure;
 890
 891         if ((key->basic.n_proto == htons(ETH_P_IP) ||
 892              key->basic.n_proto == htons(ETH_P_IPV6)) &&
 893             fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
 894                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
 895                             sizeof(key->basic.ip_proto)))
 896                 goto nla_put_failure;
 897
 898         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
 899             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
 900                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
 901                              sizeof(key->ipv4.src)) ||
 902              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
 903                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
 904                              sizeof(key->ipv4.dst))))
 905                 goto nla_put_failure;
 906         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
 907                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
 908                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
 909                                   sizeof(key->ipv6.src)) ||
 910                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
 911                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
 912                                   sizeof(key->ipv6.dst))))
 913                 goto nla_put_failure;
 914
 915         if (key->basic.ip_proto == IPPROTO_TCP &&
 916             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
 917                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
 918                              sizeof(key->tp.src)) ||
 919              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
 920                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
 921                              sizeof(key->tp.dst))))
 922                 goto nla_put_failure;
 923         else if (key->basic.ip_proto == IPPROTO_UDP &&
 924                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
 925                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
 926                                   sizeof(key->tp.src)) ||
 927                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
 928                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
 929                                   sizeof(key->tp.dst))))
 930                 goto nla_put_failure;
 931
 932         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
 933             (fl_dump_key_val(skb, &key->enc_ipv4.src,
 934                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
 935                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
 936                             sizeof(key->enc_ipv4.src)) ||
 937              fl_dump_key_val(skb, &key->enc_ipv4.dst,
 938                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
 939                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
 940                              sizeof(key->enc_ipv4.dst))))
 941                 goto nla_put_failure;
 942         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
 943                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
 944                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
 945                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
 946                             sizeof(key->enc_ipv6.src)) ||
 947                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
 948                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
 949                                  &mask->enc_ipv6.dst,
 950                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
 951                             sizeof(key->enc_ipv6.dst))))
 952                 goto nla_put_failure;
 953
 954         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
 955                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
 956                             sizeof(key->enc_key_id)))
 957                 goto nla_put_failure;
 958
 959         nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags);
 960
 961         if (tcf_exts_dump(skb, &f->exts))
 962                 goto nla_put_failure;
 963
 964         nla_nest_end(skb, nest);
 965
 966         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
 967                 goto nla_put_failure;
 968
 969         return skb->len;
 970
 971 nla_put_failure:
 972         nla_nest_cancel(skb, nest);
 973         return -1;
 974 }
 975
 976 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
 977         .kind           = "flower",
 978         .classify       = fl_classify,
 979         .init           = fl_init,
 980         .destroy        = fl_destroy,
 981         .get            = fl_get,
 982         .change         = fl_change,
 983         .delete         = fl_delete,
 984         .walk           = fl_walk,
 985         .dump           = fl_dump,
 986         .owner          = THIS_MODULE,
 987 };
 988
 989 static int __init cls_fl_init(void)
 990 {
 991         return register_tcf_proto_ops(&cls_fl_ops);
 992 }
 993
 994 static void __exit cls_fl_exit(void)
 995 {
 996         unregister_tcf_proto_ops(&cls_fl_ops);
 997 }
 998
 999 module_init(cls_fl_init);
1000 module_exit(cls_fl_exit);
1001
1002 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
1003 MODULE_DESCRIPTION("Flower classifier");
1004 MODULE_LICENSE("GPL v2");