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