net/sched/cls_flow.c

   1 /*
   2  * net/sched/cls_flow.c         Generic flow classifier
   3  *
   4  * Copyright (c) 2007, 2008 Patrick McHardy <kaber@trash.net>
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version 2
   9  * of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/init.h>
  14 #include <linux/list.h>
  15 #include <linux/jhash.h>
  16 #include <linux/random.h>
  17 #include <linux/pkt_cls.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/in.h>
  20 #include <linux/ip.h>
  21 #include <linux/ipv6.h>
  22 #include <linux/if_vlan.h>
  23 #include <linux/slab.h>
  24 #include <linux/module.h>
  25 #include <net/inet_sock.h>
  26
  27 #include <net/pkt_cls.h>
  28 #include <net/ip.h>
  29 #include <net/route.h>
  30 #include <net/flow_dissector.h>
  31
  32 #if IS_ENABLED(CONFIG_NF_CONNTRACK)
  33 #include <net/netfilter/nf_conntrack.h>
  34 #endif
  35
  36 struct flow_head {
  37         struct list_head        filters;
  38         struct rcu_head         rcu;
  39 };
  40
  41 struct flow_filter {
  42         struct list_head        list;
  43         struct tcf_exts         exts;
  44         struct tcf_ematch_tree  ematches;
  45         struct tcf_proto        *tp;
  46         struct timer_list       perturb_timer;
  47         u32                     perturb_period;
  48         u32                     handle;
  49
  50         u32                     nkeys;
  51         u32                     keymask;
  52         u32                     mode;
  53         u32                     mask;
  54         u32                     xor;
  55         u32                     rshift;
  56         u32                     addend;
  57         u32                     divisor;
  58         u32                     baseclass;
  59         u32                     hashrnd;
  60         union {
  61                 struct work_struct      work;
  62                 struct rcu_head         rcu;
  63         };
  64 };
  65
  66 static inline u32 addr_fold(void *addr)
  67 {
  68         unsigned long a = (unsigned long)addr;
  69
  70         return (a & 0xFFFFFFFF) ^ (BITS_PER_LONG > 32 ? a >> 32 : 0);
  71 }
  72
  73 static u32 flow_get_src(const struct sk_buff *skb, const struct flow_keys *flow)
  74 {
  75         __be32 src = flow_get_u32_src(flow);
  76
  77         if (src)
  78                 return ntohl(src);
  79
  80         return addr_fold(skb->sk);
  81 }
  82
  83 static u32 flow_get_dst(const struct sk_buff *skb, const struct flow_keys *flow)
  84 {
  85         __be32 dst = flow_get_u32_dst(flow);
  86
  87         if (dst)
  88                 return ntohl(dst);
  89
  90         return addr_fold(skb_dst(skb)) ^ (__force u16) tc_skb_protocol(skb);
  91 }
  92
  93 static u32 flow_get_proto(const struct sk_buff *skb,
  94                           const struct flow_keys *flow)
  95 {
  96         return flow->basic.ip_proto;
  97 }
  98
  99 static u32 flow_get_proto_src(const struct sk_buff *skb,
 100                               const struct flow_keys *flow)
 101 {
 102         if (flow->ports.ports)
 103                 return ntohs(flow->ports.src);
 104
 105         return addr_fold(skb->sk);
 106 }
 107
 108 static u32 flow_get_proto_dst(const struct sk_buff *skb,
 109                               const struct flow_keys *flow)
 110 {
 111         if (flow->ports.ports)
 112                 return ntohs(flow->ports.dst);
 113
 114         return addr_fold(skb_dst(skb)) ^ (__force u16) tc_skb_protocol(skb);
 115 }
 116
 117 static u32 flow_get_iif(const struct sk_buff *skb)
 118 {
 119         return skb->skb_iif;
 120 }
 121
 122 static u32 flow_get_priority(const struct sk_buff *skb)
 123 {
 124         return skb->priority;
 125 }
 126
 127 static u32 flow_get_mark(const struct sk_buff *skb)
 128 {
 129         return skb->mark;
 130 }
 131
 132 static u32 flow_get_nfct(const struct sk_buff *skb)
 133 {
 134 #if IS_ENABLED(CONFIG_NF_CONNTRACK)
 135         return addr_fold(skb_nfct(skb));
 136 #else
 137         return 0;
 138 #endif
 139 }
 140
 141 #if IS_ENABLED(CONFIG_NF_CONNTRACK)
 142 #define CTTUPLE(skb, member)                                            \
 143 ({                                                                      \
 144         enum ip_conntrack_info ctinfo;                                  \
 145         const struct nf_conn *ct = nf_ct_get(skb, &ctinfo);             \
 146         if (ct == NULL)                                                 \
 147                 goto fallback;                                          \
 148         ct->tuplehash[CTINFO2DIR(ctinfo)].tuple.member;                 \
 149 })
 150 #else
 151 #define CTTUPLE(skb, member)                                            \
 152 ({                                                                      \
 153         goto fallback;                                                  \
 154         0;                                                              \
 155 })
 156 #endif
 157
 158 static u32 flow_get_nfct_src(const struct sk_buff *skb,
 159                              const struct flow_keys *flow)
 160 {
 161         switch (tc_skb_protocol(skb)) {
 162         case htons(ETH_P_IP):
 163                 return ntohl(CTTUPLE(skb, src.u3.ip));
 164         case htons(ETH_P_IPV6):
 165                 return ntohl(CTTUPLE(skb, src.u3.ip6[3]));
 166         }
 167 fallback:
 168         return flow_get_src(skb, flow);
 169 }
 170
 171 static u32 flow_get_nfct_dst(const struct sk_buff *skb,
 172                              const struct flow_keys *flow)
 173 {
 174         switch (tc_skb_protocol(skb)) {
 175         case htons(ETH_P_IP):
 176                 return ntohl(CTTUPLE(skb, dst.u3.ip));
 177         case htons(ETH_P_IPV6):
 178                 return ntohl(CTTUPLE(skb, dst.u3.ip6[3]));
 179         }
 180 fallback:
 181         return flow_get_dst(skb, flow);
 182 }
 183
 184 static u32 flow_get_nfct_proto_src(const struct sk_buff *skb,
 185                                    const struct flow_keys *flow)
 186 {
 187         return ntohs(CTTUPLE(skb, src.u.all));
 188 fallback:
 189         return flow_get_proto_src(skb, flow);
 190 }
 191
 192 static u32 flow_get_nfct_proto_dst(const struct sk_buff *skb,
 193                                    const struct flow_keys *flow)
 194 {
 195         return ntohs(CTTUPLE(skb, dst.u.all));
 196 fallback:
 197         return flow_get_proto_dst(skb, flow);
 198 }
 199
 200 static u32 flow_get_rtclassid(const struct sk_buff *skb)
 201 {
 202 #ifdef CONFIG_IP_ROUTE_CLASSID
 203         if (skb_dst(skb))
 204                 return skb_dst(skb)->tclassid;
 205 #endif
 206         return 0;
 207 }
 208
 209 static u32 flow_get_skuid(const struct sk_buff *skb)
 210 {
 211         struct sock *sk = skb_to_full_sk(skb);
 212
 213         if (sk && sk->sk_socket && sk->sk_socket->file) {
 214                 kuid_t skuid = sk->sk_socket->file->f_cred->fsuid;
 215
 216                 return from_kuid(&init_user_ns, skuid);
 217         }
 218         return 0;
 219 }
 220
 221 static u32 flow_get_skgid(const struct sk_buff *skb)
 222 {
 223         struct sock *sk = skb_to_full_sk(skb);
 224
 225         if (sk && sk->sk_socket && sk->sk_socket->file) {
 226                 kgid_t skgid = sk->sk_socket->file->f_cred->fsgid;
 227
 228                 return from_kgid(&init_user_ns, skgid);
 229         }
 230         return 0;
 231 }
 232
 233 static u32 flow_get_vlan_tag(const struct sk_buff *skb)
 234 {
 235         u16 uninitialized_var(tag);
 236
 237         if (vlan_get_tag(skb, &tag) < 0)
 238                 return 0;
 239         return tag & VLAN_VID_MASK;
 240 }
 241
 242 static u32 flow_get_rxhash(struct sk_buff *skb)
 243 {
 244         return skb_get_hash(skb);
 245 }
 246
 247 static u32 flow_key_get(struct sk_buff *skb, int key, struct flow_keys *flow)
 248 {
 249         switch (key) {
 250         case FLOW_KEY_SRC:
 251                 return flow_get_src(skb, flow);
 252         case FLOW_KEY_DST:
 253                 return flow_get_dst(skb, flow);
 254         case FLOW_KEY_PROTO:
 255                 return flow_get_proto(skb, flow);
 256         case FLOW_KEY_PROTO_SRC:
 257                 return flow_get_proto_src(skb, flow);
 258         case FLOW_KEY_PROTO_DST:
 259                 return flow_get_proto_dst(skb, flow);
 260         case FLOW_KEY_IIF:
 261                 return flow_get_iif(skb);
 262         case FLOW_KEY_PRIORITY:
 263                 return flow_get_priority(skb);
 264         case FLOW_KEY_MARK:
 265                 return flow_get_mark(skb);
 266         case FLOW_KEY_NFCT:
 267                 return flow_get_nfct(skb);
 268         case FLOW_KEY_NFCT_SRC:
 269                 return flow_get_nfct_src(skb, flow);
 270         case FLOW_KEY_NFCT_DST:
 271                 return flow_get_nfct_dst(skb, flow);
 272         case FLOW_KEY_NFCT_PROTO_SRC:
 273                 return flow_get_nfct_proto_src(skb, flow);
 274         case FLOW_KEY_NFCT_PROTO_DST:
 275                 return flow_get_nfct_proto_dst(skb, flow);
 276         case FLOW_KEY_RTCLASSID:
 277                 return flow_get_rtclassid(skb);
 278         case FLOW_KEY_SKUID:
 279                 return flow_get_skuid(skb);
 280         case FLOW_KEY_SKGID:
 281                 return flow_get_skgid(skb);
 282         case FLOW_KEY_VLAN_TAG:
 283                 return flow_get_vlan_tag(skb);
 284         case FLOW_KEY_RXHASH:
 285                 return flow_get_rxhash(skb);
 286         default:
 287                 WARN_ON(1);
 288                 return 0;
 289         }
 290 }
 291
 292 #define FLOW_KEYS_NEEDED ((1 << FLOW_KEY_SRC) |                 \
 293                           (1 << FLOW_KEY_DST) |                 \
 294                           (1 << FLOW_KEY_PROTO) |               \
 295                           (1 << FLOW_KEY_PROTO_SRC) |           \
 296                           (1 << FLOW_KEY_PROTO_DST) |           \
 297                           (1 << FLOW_KEY_NFCT_SRC) |            \
 298                           (1 << FLOW_KEY_NFCT_DST) |            \
 299                           (1 << FLOW_KEY_NFCT_PROTO_SRC) |      \
 300                           (1 << FLOW_KEY_NFCT_PROTO_DST))
 301
 302 static int flow_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 303                          struct tcf_result *res)
 304 {
 305         struct flow_head *head = rcu_dereference_bh(tp->root);
 306         struct flow_filter *f;
 307         u32 keymask;
 308         u32 classid;
 309         unsigned int n, key;
 310         int r;
 311
 312         list_for_each_entry_rcu(f, &head->filters, list) {
 313                 u32 keys[FLOW_KEY_MAX + 1];
 314                 struct flow_keys flow_keys;
 315
 316                 if (!tcf_em_tree_match(skb, &f->ematches, NULL))
 317                         continue;
 318
 319                 keymask = f->keymask;
 320                 if (keymask & FLOW_KEYS_NEEDED)
 321                         skb_flow_dissect_flow_keys(skb, &flow_keys, 0);
 322
 323                 for (n = 0; n < f->nkeys; n++) {
 324                         key = ffs(keymask) - 1;
 325                         keymask &= ~(1 << key);
 326                         keys[n] = flow_key_get(skb, key, &flow_keys);
 327                 }
 328
 329                 if (f->mode == FLOW_MODE_HASH)
 330                         classid = jhash2(keys, f->nkeys, f->hashrnd);
 331                 else {
 332                         classid = keys[0];
 333                         classid = (classid & f->mask) ^ f->xor;
 334                         classid = (classid >> f->rshift) + f->addend;
 335                 }
 336
 337                 if (f->divisor)
 338                         classid %= f->divisor;
 339
 340                 res->class   = 0;
 341                 res->classid = TC_H_MAKE(f->baseclass, f->baseclass + classid);
 342
 343                 r = tcf_exts_exec(skb, &f->exts, res);
 344                 if (r < 0)
 345                         continue;
 346                 return r;
 347         }
 348         return -1;
 349 }
 350
 351 static void flow_perturbation(struct timer_list *t)
 352 {
 353         struct flow_filter *f = from_timer(f, t, perturb_timer);
 354
 355         get_random_bytes(&f->hashrnd, 4);
 356         if (f->perturb_period)
 357                 mod_timer(&f->perturb_timer, jiffies + f->perturb_period);
 358 }
 359
 360 static const struct nla_policy flow_policy[TCA_FLOW_MAX + 1] = {
 361         [TCA_FLOW_KEYS]         = { .type = NLA_U32 },
 362         [TCA_FLOW_MODE]         = { .type = NLA_U32 },
 363         [TCA_FLOW_BASECLASS]    = { .type = NLA_U32 },
 364         [TCA_FLOW_RSHIFT]       = { .type = NLA_U32 },
 365         [TCA_FLOW_ADDEND]       = { .type = NLA_U32 },
 366         [TCA_FLOW_MASK]         = { .type = NLA_U32 },
 367         [TCA_FLOW_XOR]          = { .type = NLA_U32 },
 368         [TCA_FLOW_DIVISOR]      = { .type = NLA_U32 },
 369         [TCA_FLOW_ACT]          = { .type = NLA_NESTED },
 370         [TCA_FLOW_POLICE]       = { .type = NLA_NESTED },
 371         [TCA_FLOW_EMATCHES]     = { .type = NLA_NESTED },
 372         [TCA_FLOW_PERTURB]      = { .type = NLA_U32 },
 373 };
 374
 375 static void __flow_destroy_filter(struct flow_filter *f)
 376 {
 377         del_timer_sync(&f->perturb_timer);
 378         tcf_exts_destroy(&f->exts);
 379         tcf_em_tree_destroy(&f->ematches);
 380         tcf_exts_put_net(&f->exts);
 381         kfree(f);
 382 }
 383
 384 static void flow_destroy_filter_work(struct work_struct *work)
 385 {
 386         struct flow_filter *f = container_of(work, struct flow_filter, work);
 387
 388         rtnl_lock();
 389         __flow_destroy_filter(f);
 390         rtnl_unlock();
 391 }
 392
 393 static void flow_destroy_filter(struct rcu_head *head)
 394 {
 395         struct flow_filter *f = container_of(head, struct flow_filter, rcu);
 396
 397         INIT_WORK(&f->work, flow_destroy_filter_work);
 398         tcf_queue_work(&f->work);
 399 }
 400
 401 static int flow_change(struct net *net, struct sk_buff *in_skb,
 402                        struct tcf_proto *tp, unsigned long base,
 403                        u32 handle, struct nlattr **tca,
 404                        void **arg, bool ovr)
 405 {
 406         struct flow_head *head = rtnl_dereference(tp->root);
 407         struct flow_filter *fold, *fnew;
 408         struct nlattr *opt = tca[TCA_OPTIONS];
 409         struct nlattr *tb[TCA_FLOW_MAX + 1];
 410         unsigned int nkeys = 0;
 411         unsigned int perturb_period = 0;
 412         u32 baseclass = 0;
 413         u32 keymask = 0;
 414         u32 mode;
 415         int err;
 416
 417         if (opt == NULL)
 418                 return -EINVAL;
 419
 420         err = nla_parse_nested(tb, TCA_FLOW_MAX, opt, flow_policy, NULL);
 421         if (err < 0)
 422                 return err;
 423
 424         if (tb[TCA_FLOW_BASECLASS]) {
 425                 baseclass = nla_get_u32(tb[TCA_FLOW_BASECLASS]);
 426                 if (TC_H_MIN(baseclass) == 0)
 427                         return -EINVAL;
 428         }
 429
 430         if (tb[TCA_FLOW_KEYS]) {
 431                 keymask = nla_get_u32(tb[TCA_FLOW_KEYS]);
 432
 433                 nkeys = hweight32(keymask);
 434                 if (nkeys == 0)
 435                         return -EINVAL;
 436
 437                 if (fls(keymask) - 1 > FLOW_KEY_MAX)
 438                         return -EOPNOTSUPP;
 439
 440                 if ((keymask & (FLOW_KEY_SKUID|FLOW_KEY_SKGID)) &&
 441                     sk_user_ns(NETLINK_CB(in_skb).sk) != &init_user_ns)
 442                         return -EOPNOTSUPP;
 443         }
 444
 445         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
 446         if (!fnew)
 447                 return -ENOBUFS;
 448
 449         err = tcf_em_tree_validate(tp, tb[TCA_FLOW_EMATCHES], &fnew->ematches);
 450         if (err < 0)
 451                 goto err1;
 452
 453         err = tcf_exts_init(&fnew->exts, TCA_FLOW_ACT, TCA_FLOW_POLICE);
 454         if (err < 0)
 455                 goto err2;
 456
 457         err = tcf_exts_validate(net, tp, tb, tca[TCA_RATE], &fnew->exts, ovr);
 458         if (err < 0)
 459                 goto err2;
 460
 461         fold = *arg;
 462         if (fold) {
 463                 err = -EINVAL;
 464                 if (fold->handle != handle && handle)
 465                         goto err2;
 466
 467                 /* Copy fold into fnew */
 468                 fnew->tp = fold->tp;
 469                 fnew->handle = fold->handle;
 470                 fnew->nkeys = fold->nkeys;
 471                 fnew->keymask = fold->keymask;
 472                 fnew->mode = fold->mode;
 473                 fnew->mask = fold->mask;
 474                 fnew->xor = fold->xor;
 475                 fnew->rshift = fold->rshift;
 476                 fnew->addend = fold->addend;
 477                 fnew->divisor = fold->divisor;
 478                 fnew->baseclass = fold->baseclass;
 479                 fnew->hashrnd = fold->hashrnd;
 480
 481                 mode = fold->mode;
 482                 if (tb[TCA_FLOW_MODE])
 483                         mode = nla_get_u32(tb[TCA_FLOW_MODE]);
 484                 if (mode != FLOW_MODE_HASH && nkeys > 1)
 485                         goto err2;
 486
 487                 if (mode == FLOW_MODE_HASH)
 488                         perturb_period = fold->perturb_period;
 489                 if (tb[TCA_FLOW_PERTURB]) {
 490                         if (mode != FLOW_MODE_HASH)
 491                                 goto err2;
 492                         perturb_period = nla_get_u32(tb[TCA_FLOW_PERTURB]) * HZ;
 493                 }
 494         } else {
 495                 err = -EINVAL;
 496                 if (!handle)
 497                         goto err2;
 498                 if (!tb[TCA_FLOW_KEYS])
 499                         goto err2;
 500
 501                 mode = FLOW_MODE_MAP;
 502                 if (tb[TCA_FLOW_MODE])
 503                         mode = nla_get_u32(tb[TCA_FLOW_MODE]);
 504                 if (mode != FLOW_MODE_HASH && nkeys > 1)
 505                         goto err2;
 506
 507                 if (tb[TCA_FLOW_PERTURB]) {
 508                         if (mode != FLOW_MODE_HASH)
 509                                 goto err2;
 510                         perturb_period = nla_get_u32(tb[TCA_FLOW_PERTURB]) * HZ;
 511                 }
 512
 513                 if (TC_H_MAJ(baseclass) == 0) {
 514                         struct Qdisc *q = tcf_block_q(tp->chain->block);
 515
 516                         baseclass = TC_H_MAKE(q->handle, baseclass);
 517                 }
 518                 if (TC_H_MIN(baseclass) == 0)
 519                         baseclass = TC_H_MAKE(baseclass, 1);
 520
 521                 fnew->handle = handle;
 522                 fnew->mask  = ~0U;
 523                 fnew->tp = tp;
 524                 get_random_bytes(&fnew->hashrnd, 4);
 525         }
 526
 527         timer_setup(&fnew->perturb_timer, flow_perturbation, TIMER_DEFERRABLE);
 528
 529         netif_keep_dst(qdisc_dev(tp->q));
 530
 531         if (tb[TCA_FLOW_KEYS]) {
 532                 fnew->keymask = keymask;
 533                 fnew->nkeys   = nkeys;
 534         }
 535
 536         fnew->mode = mode;
 537
 538         if (tb[TCA_FLOW_MASK])
 539                 fnew->mask = nla_get_u32(tb[TCA_FLOW_MASK]);
 540         if (tb[TCA_FLOW_XOR])
 541                 fnew->xor = nla_get_u32(tb[TCA_FLOW_XOR]);
 542         if (tb[TCA_FLOW_RSHIFT])
 543                 fnew->rshift = nla_get_u32(tb[TCA_FLOW_RSHIFT]);
 544         if (tb[TCA_FLOW_ADDEND])
 545                 fnew->addend = nla_get_u32(tb[TCA_FLOW_ADDEND]);
 546
 547         if (tb[TCA_FLOW_DIVISOR])
 548                 fnew->divisor = nla_get_u32(tb[TCA_FLOW_DIVISOR]);
 549         if (baseclass)
 550                 fnew->baseclass = baseclass;
 551
 552         fnew->perturb_period = perturb_period;
 553         if (perturb_period)
 554                 mod_timer(&fnew->perturb_timer, jiffies + perturb_period);
 555
 556         if (!*arg)
 557                 list_add_tail_rcu(&fnew->list, &head->filters);
 558         else
 559                 list_replace_rcu(&fold->list, &fnew->list);
 560
 561         *arg = fnew;
 562
 563         if (fold) {
 564                 tcf_exts_get_net(&fold->exts);
 565                 call_rcu(&fold->rcu, flow_destroy_filter);
 566         }
 567         return 0;
 568
 569 err2:
 570         tcf_exts_destroy(&fnew->exts);
 571         tcf_em_tree_destroy(&fnew->ematches);
 572 err1:
 573         kfree(fnew);
 574         return err;
 575 }
 576
 577 static int flow_delete(struct tcf_proto *tp, void *arg, bool *last)
 578 {
 579         struct flow_head *head = rtnl_dereference(tp->root);
 580         struct flow_filter *f = arg;
 581
 582         list_del_rcu(&f->list);
 583         tcf_exts_get_net(&f->exts);
 584         call_rcu(&f->rcu, flow_destroy_filter);
 585         *last = list_empty(&head->filters);
 586         return 0;
 587 }
 588
 589 static int flow_init(struct tcf_proto *tp)
 590 {
 591         struct flow_head *head;
 592
 593         head = kzalloc(sizeof(*head), GFP_KERNEL);
 594         if (head == NULL)
 595                 return -ENOBUFS;
 596         INIT_LIST_HEAD(&head->filters);
 597         rcu_assign_pointer(tp->root, head);
 598         return 0;
 599 }
 600
 601 static void flow_destroy(struct tcf_proto *tp)
 602 {
 603         struct flow_head *head = rtnl_dereference(tp->root);
 604         struct flow_filter *f, *next;
 605
 606         list_for_each_entry_safe(f, next, &head->filters, list) {
 607                 list_del_rcu(&f->list);
 608                 if (tcf_exts_get_net(&f->exts))
 609                         call_rcu(&f->rcu, flow_destroy_filter);
 610                 else
 611                         __flow_destroy_filter(f);
 612         }
 613         kfree_rcu(head, rcu);
 614 }
 615
 616 static void *flow_get(struct tcf_proto *tp, u32 handle)
 617 {
 618         struct flow_head *head = rtnl_dereference(tp->root);
 619         struct flow_filter *f;
 620
 621         list_for_each_entry(f, &head->filters, list)
 622                 if (f->handle == handle)
 623                         return f;
 624         return NULL;
 625 }
 626
 627 static int flow_dump(struct net *net, struct tcf_proto *tp, void *fh,
 628                      struct sk_buff *skb, struct tcmsg *t)
 629 {
 630         struct flow_filter *f = fh;
 631         struct nlattr *nest;
 632
 633         if (f == NULL)
 634                 return skb->len;
 635
 636         t->tcm_handle = f->handle;
 637
 638         nest = nla_nest_start(skb, TCA_OPTIONS);
 639         if (nest == NULL)
 640                 goto nla_put_failure;
 641
 642         if (nla_put_u32(skb, TCA_FLOW_KEYS, f->keymask) ||
 643             nla_put_u32(skb, TCA_FLOW_MODE, f->mode))
 644                 goto nla_put_failure;
 645
 646         if (f->mask != ~0 || f->xor != 0) {
 647                 if (nla_put_u32(skb, TCA_FLOW_MASK, f->mask) ||
 648                     nla_put_u32(skb, TCA_FLOW_XOR, f->xor))
 649                         goto nla_put_failure;
 650         }
 651         if (f->rshift &&
 652             nla_put_u32(skb, TCA_FLOW_RSHIFT, f->rshift))
 653                 goto nla_put_failure;
 654         if (f->addend &&
 655             nla_put_u32(skb, TCA_FLOW_ADDEND, f->addend))
 656                 goto nla_put_failure;
 657
 658         if (f->divisor &&
 659             nla_put_u32(skb, TCA_FLOW_DIVISOR, f->divisor))
 660                 goto nla_put_failure;
 661         if (f->baseclass &&
 662             nla_put_u32(skb, TCA_FLOW_BASECLASS, f->baseclass))
 663                 goto nla_put_failure;
 664
 665         if (f->perturb_period &&
 666             nla_put_u32(skb, TCA_FLOW_PERTURB, f->perturb_period / HZ))
 667                 goto nla_put_failure;
 668
 669         if (tcf_exts_dump(skb, &f->exts) < 0)
 670                 goto nla_put_failure;
 671 #ifdef CONFIG_NET_EMATCH
 672         if (f->ematches.hdr.nmatches &&
 673             tcf_em_tree_dump(skb, &f->ematches, TCA_FLOW_EMATCHES) < 0)
 674                 goto nla_put_failure;
 675 #endif
 676         nla_nest_end(skb, nest);
 677
 678         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
 679                 goto nla_put_failure;
 680
 681         return skb->len;
 682
 683 nla_put_failure:
 684         nla_nest_cancel(skb, nest);
 685         return -1;
 686 }
 687
 688 static void flow_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 689 {
 690         struct flow_head *head = rtnl_dereference(tp->root);
 691         struct flow_filter *f;
 692
 693         list_for_each_entry(f, &head->filters, list) {
 694                 if (arg->count < arg->skip)
 695                         goto skip;
 696                 if (arg->fn(tp, f, arg) < 0) {
 697                         arg->stop = 1;
 698                         break;
 699                 }
 700 skip:
 701                 arg->count++;
 702         }
 703 }
 704
 705 static struct tcf_proto_ops cls_flow_ops __read_mostly = {
 706         .kind           = "flow",
 707         .classify       = flow_classify,
 708         .init           = flow_init,
 709         .destroy        = flow_destroy,
 710         .change         = flow_change,
 711         .delete         = flow_delete,
 712         .get            = flow_get,
 713         .dump           = flow_dump,
 714         .walk           = flow_walk,
 715         .owner          = THIS_MODULE,
 716 };
 717
 718 static int __init cls_flow_init(void)
 719 {
 720         return register_tcf_proto_ops(&cls_flow_ops);
 721 }
 722
 723 static void __exit cls_flow_exit(void)
 724 {
 725         unregister_tcf_proto_ops(&cls_flow_ops);
 726 }
 727
 728 module_init(cls_flow_init);
 729 module_exit(cls_flow_exit);
 730
 731 MODULE_LICENSE("GPL");
 732 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
 733 MODULE_DESCRIPTION("TC flow classifier");